According to the study abstract, plant height ranged from 90.62 cm in treatment T1 to 96.12 cm in T6, which received 0.5 t ha−1 of lime and 250 ppm GA3. Treatment T9, combining 1.0 t ha−1 of lime with 250 ppm GA3, produced the most primary branches, at 12.38, and the most secondary branches, at 34.84. It also had the highest measured leaf nitrogen concentration, 1.97%, and calcium concentration, 1.70%. These were treatment-specific observations rather than results demonstrating effects across other sites or soil conditions.

Factorial analysis of variance detected significant interactions between lime rate and GA3 concentration for plant height (p < 0.01), leaf number (p < 0.05), and leaf nutrient concentrations for nitrogen, calcium and magnesium (p < 0.05). Principal component analysis grouped the growth-promoted treatments T6 and T9 together, according to the abstract. After harvest, reported soil pH values ranged from 5.13 in T1 to 6.67 in T12, although the abstract states that further validation is required. The investigators characterized the combined treatment of 1.0 t ha−1 lime and 250 ppm GA3 as the most effective within the experiment for branching and nutrient accumulation.

The evidence comes from a single field experiment and its abstract; the supplied material does not provide a full assessment of the complete study. The authors also identify several limitations and unresolved questions. Steviol glycoside concentrations were not analyzed, so the experiment did not establish whether the treatments affected stevia’s sweet compounds. Multi-location trials are needed to test the reported responses beyond the experimental setting. The findings support further testing of lime and GA3 in acidic soils, but they do not by themselves establish commercial applicability or make a universal recommendation for stevia cultivation in northern Bangladesh.

Subject of Research: Effects of combined lime and GA3 applications on field-grown stevia growth, leaf minerals and post-harvest soil pH in northern Bangladesh

Article Title: Integration of lime and GA3 to enhance growth and leaf mineral composition of stevia (Stevia rebaudiana) in acidic soils of the northern region of Bangladesh

Article References: Amin, M., Islam, M. A., Akter, R., Mahabub, M. R., Chowdhury, M. T. I., Bhuiya, S. A., & Islam, S. (2026). Integration of lime and GA3 to enhance growth and leaf mineral composition of stevia (Stevia rebaudiana) in acidic soils of the northern region of Bangladesh. BMC Plant Biology. https://doi.org/10.1186/s12870-026-09844-5

Image Credits: AI Generated

DOI: 10.1186/s12870-026-09844-5

Keywords: stevia, acidic soils, lime, gibberellic acid, field experiment, Bangladesh, plant growth, leaf nutrients, soil pH, randomized block design, Integration, enhance

The motivation is starkly quantitative. The authors note that residents of Malaysia’s Klang Valley collectively waste more than 250 million hours each year stuck in traffic, an economic drag equivalent to between 1.1 and 2.2 percent of national GDP. Congestion also inflates vehicular emissions, placing Malaysia among the highest CO2 emitters in ASEAN. Traditional remedies such as building more roads have repeatedly run into the Downs-Thomson paradox, in which added capacity simply induces additional demand. The alternative is proactive management through Intelligent Transportation Systems, which depend fundamentally on accurate short-term traffic forecasting to enable dynamic routing, variable toll pricing, and efficient resource allocation.

Deep learning has become the dominant approach to that forecasting problem, and for good reason. Long Short-Term Memory networks, designed to preserve information across long sequences without the vanishing gradient problems that plague ordinary recurrent networks, excel at capturing the daily and weekly rhythms of traffic flow. But the Malaysian team identified a subtle weakness: standalone recurrent architectures struggle to model the instantaneous cross-correlations between heterogeneous vehicle classes, such as the synchronized surges of passenger cars and heavy commercial vehicles within a single time window. Their solution was to prepend a one-dimensional convolutional layer that acts as a trainable spatial filter, extracting localized inter-class dependencies before the temporal memory cells process the sequence.

The hybrid CNN-LSTM engine was validated against an unusually rich dataset: five years of multivariate toll records from PLUS Malaysia Berhad covering 18 plazas in the Klang Valley, totaling 73,891 aggregated observations across five vehicle classes. The data was normalized with min-max scaling and transformed into a supervised learning problem using a 24-hour sliding window, with a strict 70:15:15 chronological split to prevent data leakage. The convolutional head compresses the raw multivariate input into latent spatial features, the LSTM encodes their temporal evolution, and a final ensemble layer built on XGBoost or Random Forest refines the decision boundary using the latent vector produced by the neural backbone.

The benchmarking results were decisive. The CNN-LSTM consistently outperformed standalone LSTM, Bidirectional LSTM, and GRU baselines, achieving the lowest Root Mean Squared Error and Mean Absolute Percentage Error across the heavy commercial and specialty vehicle classes. The contrast was most dramatic for Class 5 machinery, where the CNN-LSTM posted an RMSE of roughly 744 while the GRU’s error ballooned to nearly 2,071, with a MAPE above 52 percent. Standalone recurrent models, the authors argue, simply cannot cope with the sporadic, non-uniform distributions of heavy machinery and buses, whereas the convolutional front end maps localized spikes before they reach the memory cells, keeping forecasts within a tighter band.

Accuracy alone, however, was never the point. The study’s central innovation is its treatment of explainability as a triangulated, multi-perspective exercise rather than a single post-hoc calculation. Most transportation research relies on SHAP, a game-theoretic method that assigns each feature a marginal contribution to a prediction, but the authors argue that using SHAP in isolation leaves the system vulnerable to that method’s inherent biases. Their framework instead cross-validates three mathematically independent explainers: SHAP, Integrated Gradients, which integrates gradients along a path from a baseline input to the actual input, and LIME, which fits a local linear surrogate around each prediction. Only when these methods agree can the attribution be trusted.

The triangulation paid off in a striking way. Both SHAP and Integrated Gradients overwhelmingly identified Class 1 passenger vehicles and total volume as the dominant drivers of the model’s internal state, with normalized importance scores between 0.98 and 1.00, while LIME alone would have suggested a far more distributed picture. That consensus then served as the bridge to the symbolic layer. The team discretized the continuous attribution scores into a binary formal context, using median-based thresholds to avoid both over-saturation and sparsity, and applied Formal Concept Analysis to extract a concept lattice, a hierarchical structure in which each node represents a maximal set of traffic observations sharing a common set of attributes.

The lattice yields deterministic implication rules with remarkable confidence levels. One rule states that when passenger cars, large trucks, buses, and total volume simultaneously exceed their operational thresholds at a given toll plaza, a spike in small lorries follows with absolute certainty. Another provides the reciprocal logic: surges in heavy commercial traffic inherently signal corresponding surges in passenger car volume. Because these symbolic rules align with the numerical consensus from SHAP and Integrated Gradients, the authors can demonstrate what they call neuro-symbolic alignment: the neural network has genuinely learned to base its predictions on the same foundational variables that govern the human-readable logic, rather than exploiting spurious correlations that no one can audit.

The practical implications extend beyond academic curiosity. The full pipeline, from convolutional filtering through ensemble regression to multi-perspective attribution extraction, achieves an average inference latency of 14.2 milliseconds per 24-hour look-back window on an NVIDIA RTX 4090, comfortably within the 15-minute to 1-hour cycles typical of dynamic traffic management. For highway operators, that means forecasts can be paired with explicit rules for proactive intervention, such as anticipating gridlock before it manifests, and that algorithmic actions in automated tolling or lane allocation can always be backed by a symbolic reason. The authors also argue the architecture is network-agnostic: because it relies on multivariate sequences rather than rigid graph topologies, adapting it to a new corridor requires only retraining the feature extractor on local sensor streams.

The study, published as open access in Applied Intelligence, positions itself as a benchmark for deploying explainable AI in high-stakes transportation infrastructure, and it arrives amid a broader shift in the field toward neuro-symbolic systems that pair connectionist learning with formal logic. The authors acknowledge limitations and chart future work in three directions: real-time integration into live traffic control systems, extension to multimodal networks encompassing bus rapid transit and rail flows, and the development of entropy-driven adaptive thresholding so that the concept lattices can continuously evolve with seasonal traffic regimes. For now, the Malaysian dataset demonstrates the core claim: a forecasting system can be simultaneously neural in its handling of noisy real-world data and symbolic in the logic it offers to the humans who must act on it, closing the trust gap that has long separated deep learning from operational deployment.

Subject of Research: Neuro-symbolic explainable AI framework combining CNN-LSTM deep learning and Formal Concept Analysis for interpretable traffic volume forecasting

Article Title: Multi-Perspective explainability in traffic prediction: A neuro-symbolic approach using CNN-LSTMs and concept lattices

Article References: Mohamad Yusof, S. S., Ab. Ghani, H., Md Saleh, N. I., & Khamis, N. (2026). Multi-Perspective explainability in traffic prediction: A neuro-symbolic approach using CNN-LSTMs and concept lattices. Applied Intelligence, 56(16), Article 505. https://doi.org/10.1007/s10489-026-07512-7

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07512-7

Keywords: traffic forecasting, explainable AI, CNN-LSTM, Formal Concept Analysis, neuro-symbolic AI, intelligent transportation systems, SHAP, LIME, Integrated Gradients, deep learning, Malaysia, time-series prediction

The review, which followed PRISMA 2020 guidelines and screened more than 3,300 records from six databases before including 330 studies, makes clear why the black-box problem matters so much in this field. Machine learning models can already predict diagnosis, treatment response, and relapse risk from structural and functional MRI, diffusion tensor imaging, EEG, speech, electronic health records, and genetic data. But deep neural networks that achieve high accuracy often cannot explain which features drove their decisions. In medicine, where an incorrect prediction can carry serious consequences and where regulatory frameworks such as the GDPR emphasize accountability and explainability, that opacity is a barrier to clinical adoption. XAI aims to bridge the gap between complex algorithms and human clinical reasoning by revealing influential features and decision pathways.

The technical toolkit is diverse. Model-specific techniques such as Grad-CAM generate class-specific heatmaps that highlight diagnostically important regions in medical images, with Grad-CAM++ improving localization when multiple pathological areas overlap. Model-agnostic approaches such as SHAP and LIME can explain nearly any machine learning model by quantifying the contribution of individual features, though at substantial computational cost. Layer-wise Relevance Propagation and Deep Taylor Decomposition trace predictions through layers of computation to produce detailed relevance maps, while counterfactual explanations show how small changes in input would flip a prediction, supporting a form of causal reasoning. Explanations can be global, describing model behavior across an entire dataset, or local, explaining why one particular patient was flagged as high risk. Newer directions combine XAI with self-supervised and federated learning to protect privacy, and with multimodal fusion frameworks that dynamically weight contributions from MRI, EEG, genomics, and clinical data.

In neuroimaging, the results are striking but nuanced. One deep learning model trained on grey matter volume features from 925 participants across nine international centers achieved an average AUC of 82 percent under leave-one-center-out validation, rising to 83 percent when white matter and cerebrospinal fluid features were added, with peak performance of 96 percent at the best-performing site after ComBat harmonization corrected for scanner differences. Crucially, SHAP analysis revealed that the model relied on biologically plausible regions, including the thalamus, pallidum, amygdala, hippocampus, and postcentral gyrus, findings consistent with the existing literature. Grey matter volume contributed most to classification, while the other structural measures added only marginal gains. Earlier studies reported similarly impressive numbers: a VGG16-based convolutional network with transfer learning reached 84.3 percent accuracy on fMRI data, and an ensemble framework called EMPaSchiz achieved 87 percent accuracy using resting-state fMRI from antipsychotic-naïve patients.

EEG offers a cheaper and more scalable alternative, capturing abnormal neural oscillations with millisecond temporal resolution. Traditional machine learning pipelines using wavelet decomposition, entropy measures, and connectivity features have achieved accuracies between roughly 80 and 99 percent, while deep architectures including CNNs, BiLSTM models, and graph convolutional networks have reported results above 95 percent. One study combining XGBoost, SVM, and AdaBoost with SHAP analysis reached 93 percent accuracy, with interpretable features including IQ, sex, and specific EEG wave patterns. But the review’s authors urge caution: some of the highest reported figures come from small, homogeneous datasets vulnerable to data leakage, and performance that looks implausibly perfect in cross-validation rarely survives contact with external cohorts. Meta-analytic estimates are more sobering, with pooled MRI-based classification showing sensitivity of about 79 percent and specificity of about 80 percent.

Some of the most original work moves beyond the brain scan entirely. In the Italian SPECTRA project, researchers are combining speech, emotion, and locomotion data to identify treatment-resistant schizophrenia, a condition affecting roughly 30 percent of patients that is often diagnosed too late for clozapine to be started promptly. Using multi-camera motion capture and a dual-transformer architecture to reconstruct three-dimensional skeletal movement, the team identified 16 quantifiable gait markers associated with the illness. A SHAP-based visualization renders the model’s reasoning as a color-coded matrix, with red regions marking body joints and video frames that push the prediction toward schizophrenia and blue regions marking those that pull against it, allowing clinicians to inspect exactly which movements the algorithm found telling. Features such as head posture, walking speed, arm sway, and step-length variability, drawn from prior clinical work, anchor the analysis in measurable behavior.

Language and handwriting offer another window. Formal thought disorder is a core feature of schizophrenia, and recent findings suggest that specific PANSS items such as conceptual disorganization and unusual thought content predict treatment resistance. A new 16-item screening instrument for schizophrenic features in writing, validated in 157 participants, produced markedly higher scores in patients than controls and reached 93 percent sensitivity and specificity at a cutoff of 2, with discriminant analysis classifying cases with roughly 90 percent accuracy. On the treatment side, researchers combined the Personalized Advantage Index with Bayesian Rule Lists, an XAI method that generates transparent if-then rules with explicit cutoffs. Applied to a randomized trial of paliperidone, the framework identified a treatment-indicated subgroup whose effect size reached a Cohen’s d of 1.51, far exceeding the 0.82 seen in the overall sample, while consistently naming disturbance of volition and uncooperativeness as key predictors.

Where might AI contribute most immediately? The review argues that continuous monitoring, rather than one-time diagnosis, is the field’s most realistic near-term value. Smartphone-based digital phenotyping has demonstrated that statistically significant behavioral anomalies can precede relapse, and a 2025 study using the CrossCheck dataset forecast mental states hours to days in advance from passive sensor data using XGBoost and LSTM models. Digital medicine systems have already shifted real outcomes: in a phase 3b mirror-image trial, aripiprazole tablets with ingestible sensors were associated with reduced psychiatric hospitalization and a 26.5 percentage-point improvement in medication coverage during the first three months. Yet the evidence is mixed elsewhere. The Actissist trial, a powered randomized study of a CBT-informed app for early psychosis, failed to beat digital symptom monitoring, a reminder that digital delivery alone does not guarantee clinical benefit.

The authors are unsparing about the field’s weaknesses. Risk-of-bias analysis of a representative subset found high overall risk in eight of thirteen studies, largely from overfitting and inadequate validation. Diagnostic models mostly separate patients from healthy controls, an easier problem than distinguishing schizophrenia from overlapping psychiatric syndromes. Regulatory space remains nearly empty: FDA materials note that more than 1,200 AI-enabled medical devices have been authorized, but none for mental health indications, and generative AI chatbots raise specific concerns that they may validate or amplify delusional content in vulnerable users. Patient perspectives on trust, privacy, and acceptance are almost entirely absent from the literature.

The review closes with a concrete roadmap. Within one to three years, the field needs standardized multicenter datasets and clinically meaningful explainability metrics such as explanation fidelity, stability, and actionability. Within three to five years, prospective studies should test workflow integration and clinician trust against reporting standards including CONSORT-AI and STARD-AI, while within a decade, diagnostic systems should demonstrate externally validated AUCs of at least 0.90, clinician agreement with explanations above 80 percent, a 15 to 20 percent improvement in relapse prediction, and measurable gains in patient-centered outcomes such as adherence and quality of life, all under the transparency and oversight requirements of the EU AI Act. The verdict on explainable AI in schizophrenia is clear-eyed: scientifically promising, particularly for multimodal prediction and continuous monitoring, but clinically immature. The next decade will be judged not by headline accuracy figures but by whether these transparent systems reduce relapse, improve functioning, and earn the trust of clinicians and patients alike, as an adjunct to expert psychiatric care rather than a replacement for it.

Subject of Research: Explainable artificial intelligence for transparent clinical decision-making in schizophrenia

Article Title: Explainable artificial intelligence (XAI) in schizophrenia: enhancing transparency in clinical decision-making

Article References: Srivastava, K., & Srivastava, R. (2026). Explainable artificial intelligence (XAI) in schizophrenia: enhancing transparency in clinical decision-making. Frontiers in Systems Biology, 6, Article 1898853. https://doi.org/10.3389/fsysb.2026.1898853

Image Credits: AI Generated

DOI: 10.3389/fsysb.2026.1898853

Keywords: explainable AI, schizophrenia, machine learning, SHAP, LIME, Grad-CAM, neuroimaging, EEG, digital phenotyping, precision psychiatry, treatment-resistant schizophrenia, clinical decision support

The research, led by Raimondo Fanale and colleagues at Universitas Mercatorum in Rome, working with the National Interuniversity Consortium for Telecommunications, focuses on histopathology, the microscopic examination of stained tissue sections that remains the gold standard for cancer diagnosis. The team built their system on the BreaKHis dataset, a public collection of 7,909 breast tissue micrographs from 82 patients, spanning four magnification levels and eight tumor subtypes, from benign adenosis to malignant ductal carcinoma.

At the heart of the pipeline is a knowledge-distilled convolutional neural network. An InceptionV3 model, prized for its multi-scale feature extraction, served as the teacher, while a DenseNet-121 student learned to mimic its outputs. Distillation compresses the teacher’s knowledge into a more compact student, and the payoff was striking: the student achieved a training-to-validation accuracy gap of less than 0.002, indicating negligible overfitting, while surpassing its own teacher on every test metric.

On the 1,187-image held-out test set, the distilled student reached an accuracy of 0.9916, an F1-score of 0.9939, and an area under the receiver-operating-characteristic curve of 0.9989, with sensitivity of 0.9939 and specificity of 0.9866. Five-fold cross-validation using the dataset’s official patient-level folds yielded a mean accuracy of 0.8962 and a mean AUC of 0.9598, figures the authors describe as the more conservative and representative estimate of generalization, since the single test cohort of 13 patients represents a favorable draw.

But raw accuracy is only half the story. The framework’s real innovation lies in how it interrogates the model’s reasoning. For every image, three complementary explanation techniques generate heatmaps: Grad-CAM traces gradients through the final convolutional layers to localize diagnostically relevant regions; LIME perturbs superpixels and fits a local surrogate model to measure each region’s influence; and SHAP, grounded in cooperative game theory, assigns attribution scores quantifying how much each part of the image pushes the prediction toward malignancy or benignity.

These three maps are then fused into the ExpiScore, a composite interpretability index with formal mathematical properties including boundedness, monotonicity, and linearity in its balancing parameter. The score combines a mean stability term, reflecting the average quality of the three individual explanations, with an OGSL consensus term that measures overlap, geometric coherence, and structural similarity across methods. Higher ExpiScore values signal explanations that are not only strong but mutually consistent, a property the authors argue is essential for clinical credibility.

The results reveal a striking pattern. ExpiScore is highest for true positives at 0.313, followed by true negatives at 0.279, false positives at 0.258, and false negatives at just 0.224. Correctly identified malignant cases, the most clinically critical category, consistently exhibit the most coherent explanations, while missed cancers show the weakest. A Spearman correlation of 0.674 between ExpiScore and model confidence confirms the two signals are related but far from redundant, meaning explanations carry information that confidence alone cannot capture.

This insight powers the framework’s most consequential component: a trust meta-model built from logistic regression and a decision tree, trained on 42 features derived from the explanation maps. The meta-model achieved an AUC of 0.9030 in distinguishing correct from incorrect predictions on held-out data. When deployed as a routing mechanism, it flags the 20 percent of cases with the highest predicted error probability for expert review, raising post-routing accuracy from 0.9898 to 0.9983 and cutting false negatives by 83.3 percent, under the assumption that experts correctly resolve all flagged cases.

The authors are careful to keep two operational modes strictly separate. In fully automatic mode, the CNN alone achieves the headline figures without any human involvement. In human-in-the-loop mode, the meta-model triages cases into three categories: accept for high-confidence, coherent predictions; review for ambiguous ones; and defer for cases exhibiting both low confidence and low explanatory coherence. The improvement in the second mode is decisional rather than algorithmic, as the CNN’s predictions remain unchanged while the system becomes better at identifying which ones deserve scrutiny.

Sensitivity analyses confirm the approach is robust: the balancing parameter performs at least as well at its default value as at any alternative, and even if experts resolve only 70 percent of routed cases correctly, system accuracy still exceeds the CNN-only baseline. The authors acknowledge limitations, including the computational cost of roughly 25 to 45 seconds per image for full explanation generation, the single-institution dataset, and the need for external multi-cohort validation before clinical deployment. Yet the core message stands: explainability is not merely a transparency accessory but an operational instrument that can catch the errors confidence alone conceals.

Subject of Research: A quantitative explainability framework for deep learning-based breast cancer histopathology classification

Article Title: A quantitative framework for explainability assessment in medical imaging

Article References: Fanale, R., Caldelli, R., Martini, B., & Sciarrone, F. (2026). A quantitative framework for explainability assessment in medical imaging. Frontiers in Imaging, 5, Article 1846414. https://doi.org/10.3389/fimag.2026.1846414

Image Credits: AI Generated

DOI: 10.3389/fimag.2026.1846414

Keywords: explainable AI, histopathology, breast cancer, Grad-CAM, SHAP, LIME, knowledge distillation, medical imaging, deep learning, ExpiScore, clinical decision support, false negatives

The stakes are enormous. In China alone, more than 70,000 new cases of prolonged disorders of consciousness arise annually, generating medical expenditures in the tens of billions of RMB and placing crushing burdens on families and the health system. The clinical gold standard, the Coma Recovery Scale-Revised, depends on behavioral responses and intact motor pathways, and its scoring is vulnerable to observer bias. A patient with locked-in features or severe motor impairment may retain flickers of awareness that no bedside examination can elicit. What clinicians need is an objective, physiology-based readout of the brain that does not require the patient to move, speak, or respond on command.

That is precisely what the new study, published in Applied Intelligence, set out to deliver. The researchers collected data at Luqiao Hospital of the Taizhou Enze Medical Center using functional near-infrared spectroscopy, or fNIRS, a non-invasive technique that shines near-infrared light at two wavelengths, 730 and 850 nanometers, through the scalp and measures how much is absorbed. Because oxygenated and deoxygenated hemoglobin absorb light differently, the modified Beer-Lambert law allows researchers to compute continuous changes in oxygenated hemoglobin (HbO), deoxygenated hemoglobin (HbR), and total hemoglobin (HbT) across the cortex. Compared with EEG, MEG, or fMRI, fNIRS is cheap, portable, tolerant of movement and environment, and well suited to the intensive care unit, where scanning a patient inside a giant magnet is rarely an option.

The study enrolled 17 patients in a vegetative state or unresponsive wakefulness syndrome, 22 in a minimally conscious state, and 20 healthy controls, with diagnoses independently assigned by two experienced clinicians using the CRS-R scale. Each participant lay supine wearing a cap of 13 LED emitters and 15 avalanche photodiode detectors, forming 35 measurement channels that sampled at 11 hertz and covered bilateral frontotemporal and prefrontal cortex. Rather than calling each patient’s name, which pilot testing revealed produced inconsistent responses due to dialect differences and naming habits, the team used music. Each session comprised 18 thirty-second musical excerpts, spanning piano pieces, popular songs, and saxophone solos, each followed by a 30-second rest. Music was chosen deliberately: as a non-semantic stimulus it engages auditory, emotional, memory, and attention networks regardless of language or cultural background, and it induces broader cortical activation than pure tones.

Raw optical signals were converted to optical density values, cleaned of motion artifacts using a sliding-window standard-deviation method with spline interpolation, band-pass filtered between 0.01 and 0.2 hertz with a fourth-order Butterworth filter, and transformed into hemoglobin concentrations with baseline correction against the pre-task rest. Before the model ever saw the data, the raw signals already told a striking story. Vegetative-state patients showed flat, regular traces with modest, uniform fluctuations across channels, a pattern consistent with globally suppressed neural activity. Minimally conscious patients exhibited more complex, channel-dependent dynamics, hinting that some regions were regaining responsiveness. Healthy controls displayed rich, varied, and clearly differentiated signals across all 35 channels, matching the heavy metabolic demands of a fully conscious brain.

The heart of the work is a neural architecture the authors call pDOC-STANet, a spatiotemporal attention network designed to squeeze every drop of information from those signals. Each trial contains 330 time points across 35 channels, and the three hemoglobin species are concatenated along the channel dimension, producing a 105-by-330 matrix. The temporal branch processes this matrix through an improved InceptionTime network: each Inception block runs parallel convolutions with kernel sizes of 10, 20, and 40, capturing short-term spikes, mid-range dependencies, and long-term trends simultaneously, while a residual connection guards against vanishing gradients. The authors then embedded a Squeeze-and-Excitation module in every block, which uses global average pooling and two fully connected layers to learn channel-wise attention weights, letting the network amplify the most informative temporal features automatically.

The second branch treats the same data as an image. Using the Gramian Angular Difference Field, a technique that converts a normalized time series into polar coordinates and computes pairwise angular differences between every pair of time points, each trial becomes a stack of 105 channel-wise GADF maps of size 330 by 330, a tensor of more than 11 million values. This encoding preserves global temporal relationships that a simple reshaping would destroy. A modified ResNet18, with its first convolution widened from three RGB channels to 105, digests these images end-to-end. Finally, an attention-based fusion module projects the two feature vectors into a shared latent space, concatenates them, and learns a softmax weighting over the two modalities for each sample, so the classifier can lean on whichever representation is more discriminative for a given patient.

Evaluated with a strictly subject-independent, stratified five-fold cross-validation, in which all 18 trials from any one participant stayed within a single fold to prevent data leakage, pDOC-STANet reached an accuracy of 88.72 percent, with precision of 88.50 percent and recall of 89.62 percent, on 1,062 non-overlapping samples from 59 participants. Traditional methods lagged far behind: linear discriminant analysis managed only 55.64 percent, decision trees 66.17 percent, and logistic regression, the best classical baseline, 77.44 percent. Among deep-learning approaches, InceptionTime alone achieved 84.96 percent and the GADF-ResNet branch alone 83.46 percent, both beaten by the fusion model with a statistically significant margin confirmed by paired t-testing across folds.

Ablation experiments made the architecture’s logic plain. Naively concatenating the two branches without attention actually degraded performance to 82.71 percent, below either branch alone, showing that multimodal fusion is not automatic and that redundant cross-modal noise must be filtered. Restoring attention weighting lifted accuracy to the full 88.72 percent. Replacing the GADF encoding with a direct two-dimensional reshape cost nearly two and a half points, confirming that the angular encoding preserves genuinely useful temporal dependencies. Modality experiments revealed another surprise: total hemoglobin alone, which reflects overall cerebral blood volume, was the strongest single signal at 85.71 percent accuracy, well above HbO or HbR alone, yet the three-way combination remained best, indicating that oxygenation dynamics carry complementary information. Task-state data collected during music also clearly beat resting-state recordings, which reached only 81.20 percent accuracy, evidence that musical stimulation actively elicits discriminative hemodynamic responses tied to consciousness level.

Crucially, the team did not stop at accuracy. Using LIME for class-specific local explanations and SHAP for global feature importance, they showed that the model’s decisions rest on physiologically plausible territory: healthy-control predictions leaned on HbO and HbR features in left frontotemporal and prefrontal channels, minimally conscious predictions concentrated in bilateral prefrontal regions, and vegetative-state predictions drew broadly on HbR and HbT features across both hemispheres. The researchers are candid about the limits: 59 participants from a single center cannot establish clinical generalization, labels ultimately derive from CRS-R scores, and saliency maps are feature-importance tools rather than causal neuroscience. They plan multi-center validation with more than 200 cases and a lightweight version of the model for portable embedded devices. If those efforts succeed, a bedside cap, thirty minutes of music, and an algorithm could soon give voice, in hemodynamic ink, to patients who cannot speak for themselves.

Subject of Research: Automated assessment of prolonged disorders of consciousness using fNIRS and a spatiotemporal attention deep-learning network

Article Title: pDOC-STANet: A spatiotemporal attention network for automatic assessment of prolonged disorders of consciousness using fNIRS

Article References: Li, Z., Xing, Y., Wang, S., Cui, K., Chen, X., Wu, H., Feng, Y., Yi, P., & He, S. (2026). pDOC-STANet: A spatiotemporal attention network for automatic assessment of prolonged disorders of consciousness using fNIRS. Applied Intelligence, 56(16), Article 493. https://doi.org/10.1007/s10489-026-07542-1

Image Credits: AI Generated

DOI: 10.1007/s10489-026-07542-1

Keywords: fNIRS, prolonged disorders of consciousness, deep learning, attention mechanism, InceptionTime, ResNet, Gramian Angular Difference Field, vegetative state, minimally conscious state, consciousness assessment, LIME, SHAP

The core question the researchers set out to answer is deceptively simple: can a person’s risk of smartphone addiction be predicted accurately from measurable patterns of behavior, and can the model explain exactly which behaviors drive that prediction? This second half of the question matters enormously in practice. Black-box predictive models have long been viewed with suspicion in health and social science contexts, because clinicians, educators, and policymakers are reluctant to act on a risk score they cannot interrogate. The study’s answer is that high accuracy and interpretability need not be traded against each other. After preprocessing and feature transformation of the behavioral data, the team evaluated multiple classification models, and logistic regression emerged as the strongest performer, achieving an accuracy of 0.978 and a ROC-AUC of 0.999, figures that indicate near-ceiling discrimination between addicted and non-addicted usage profiles.

The choice of logistic regression as the winning model is itself instructive. In an era when deep neural networks dominate headlines, the study demonstrates that a linear classifier, whose coefficients can be read directly as the direction and magnitude of each feature’s influence, can outperform more complex alternatives when the underlying signal in the data is strong. Logistic regression estimates the probability that a given usage profile belongs to the addicted class by fitting a weighted sum of features through a logistic function. Its transparency means that every prediction can be traced back to specific behavioral inputs, which is precisely the property that regulatory frameworks and clinical ethics increasingly demand of automated decision systems.

Statistical analysis reinforced the machine learning results in striking fashion. The researchers report a p-value of approximately 8.71 times 10 to the power of negative 93 for the effect of daily usage on addiction, an astronomically small number that indicates the relationship between daily usage and addictive outcomes is overwhelmingly significant. In practical terms, a p-value of this magnitude means that the observed association would essentially never arise by chance if no true relationship existed. Daily usage emerged as the dominant behavioral signal, a finding that aligns with a substantial body of prior literature, including concurrent studies linking usage intensity to mobile phone addiction risk and meta-analyses documenting the effects of smartphone addiction on learning outcomes.

What elevates the study beyond a standard classification exercise is its systematic application of explainable artificial intelligence, or XAI. The authors deployed four complementary techniques: SHAP, LIME, Partial Dependence Plots, and Individual Conditional Expectation curves. SHAP, which stands for SHapley Additive exPlanations, borrows a concept from cooperative game theory to distribute the credit for each prediction fairly among the input features, providing a principled account of how much each behavior pushed a given user’s risk score up or down. LIME, or Local Interpretable Model-agnostic Explanations, takes the opposite approach, fitting a simple, interpretable model in the immediate neighborhood of a single prediction so that the reasoning behind that individual classification can be inspected directly.

Partial Dependence Plots and Individual Conditional Expectation curves address the global and local structure of the model’s behavior across the feature space. A Partial Dependence Plot shows how the model’s predicted probability of addiction changes on average as one feature, such as daily screen time, varies while other features are held in check, revealing thresholds and saturation effects that a single coefficient cannot capture. Individual Conditional Expectation curves disaggregate that average picture, tracing the response for individual users so that heterogeneity in the population becomes visible. Together, these four methods give researchers, clinicians, and app designers a layered view of the model: global importance rankings, local case-level explanations, and dose-response curves that show where behavioral risk begins to climb steeply.

The significance of this methodological package extends well beyond the specific dataset. Screening instruments for smartphone addiction, such as the Smartphone Addiction Scale and the Smartphone Overuse Screening Questionnaire developed in earlier research, rely on self-reported questionnaires that are vulnerable to response bias and require willing participation. A behavioral analytics pipeline that infers risk from objective usage features could, in principle, be integrated into digital wellbeing tools, allowing at-risk users to be flagged earlier and more consistently. The authors argue that the combination of high accuracy and interpretability makes their approach appropriate for practical deployment, a claim that rests on the model’s ability to justify its predictions to the humans who would act on them.

The study also situates itself within a rapidly growing research conversation. Earlier work has applied machine learning to problematic smartphone use among children and adolescents in South Korea with explainable methods, and integrative models have incorporated personality traits alongside usage data to predict addiction in teenagers. The present study contributes to this literature by demonstrating that a deliberately simple, interpretable classifier can match or exceed the performance one might expect from more elaborate architectures, provided the behavioral features are well constructed. It also underscores a recurring theme in the field: daily usage quantity is not merely correlated with addiction but appears to be its most powerful statistical signature, dwarfing the contribution of other measured behaviors.

Cautionary notes remain. The authors state that no new datasets were generated or analyzed during the study beyond the behavioral data used, and the underlying data source is a publicly available dataset on teen smartphone usage and addiction, which means the findings are bounded by how that dataset was collected and labeled. Self-reported addiction labels, even when paired with objective usage features, carry inherent measurement uncertainty, and generalization across age groups, cultures, and device ecosystems will require further validation. The authors also note that the research involved no studies with human participants, meaning the model operates on pre-existing behavioral records rather than prospective clinical follow-up. These limitations do not diminish the central result, but they frame it: this is a methodological demonstration of what explainable behavioral analytics can achieve, not yet a validated clinical screening instrument.

Nevertheless, the implications are considerable. As smartphone manufacturers expand digital wellbeing dashboards that already log screen time, app launches, and notification counts, the raw material for addiction-risk analytics is being generated on billions of devices every day. The study’s demonstration that a transparent logistic regression model, explained through SHAP, LIME, Partial Dependence Plots, and Individual Conditional Expectation curves, can predict addiction with an accuracy of 0.978 and a ROC-AUC of 0.999 suggests that early-warning systems for problematic use are technically feasible with methods simple enough to audit. The open question, which the authors’ work now sharpens, is not whether such predictions can be made, but how societies will choose to act on them, balancing the promise of early intervention against the privacy and autonomy concerns that always accompany behavioral surveillance at scale.

Subject of Research: Explainable machine learning prediction of smartphone addiction from behavioral usage data

Article Title: Behavioral analytics for smartphone addiction prediction: an explainable and interpretable machine learning approach

Article References: Urkude, V. R., Vairachilai, S., Urkude, S. V., & Bharaneedharan, S. V. B. (2026). Behavioral analytics for smartphone addiction prediction: an explainable and interpretable machine learning approach. SN Social Sciences, 6(9), Article 410. https://doi.org/10.1007/s43545-026-01697-x

Image Credits: AI Generated

DOI: 10.1007/s43545-026-01697-x

Keywords: smartphone addiction, machine learning, explainable AI, SHAP, LIME, logistic regression, behavioral analytics, digital wellbeing, problematic smartphone use, statistical analysis, prediction models, SN Social Sciences

The study sits at the intersection of two research traditions that have been converging for years: Educational Data Mining, a field devoted to extracting useful patterns from the digital traces students leave behind, and the broader movement to apply artificial intelligence to sustainable development. The authors frame their work explicitly in terms of institutional sustainability, arguing that a university’s ability to identify and support struggling students is not merely a pastoral concern but a structural one. Student success rates and the quality of education feed directly into institutional rankings, funding decisions, and reputations, which means that a reliable early-warning system has consequences far beyond any individual classroom.

To build their model, the researchers collected primary data from 200 university students using a structured questionnaire with Likert-scale items designed to measure learning behavior, social media usage, and psychological traits, alongside academic records used strictly for research purposes in accordance with institutional ethical guidelines. What distinguishes this dataset from many earlier efforts in the field is its breadth. Rather than focusing narrowly on grades and attendance, the team captured seven categories of potential influence on academic outcomes: demographic background, socioeconomic conditions, academic history, social and family environment, health and wellness, infrastructure and services available to the student, and time management and extracurricular activity. This holistic framing reflects a growing recognition in education research that performance is shaped by a dense web of factors extending well beyond the lecture hall.

With the dataset assembled, the researchers trained and compared several regression algorithms, each representing a different mathematical philosophy for mapping input variables to a predicted outcome. The contenders included a Decision Tree Regressor, which partitions the data through a series of branching questions; a Random Forest Regressor, which averages the outputs of many decision trees to reduce overfitting; and a Support Vector Regressor, which fits a function within a tolerance margin around the training points. They also deployed Elastic Net Regression, a hybrid technique that blends two regularization penalties — the L1 penalty of lasso regression and the L2 penalty of ridge regression — into a single objective function. Regularization is the mathematical equivalent of a discipline mechanism: it constrains the model’s coefficients so that the algorithm cannot chase noise in the training data, and the elastic net’s dual penalty allows it both to shrink irrelevant features toward zero and to handle situations where predictor variables are correlated with one another, a common condition in survey data where questions about family support, finances, and wellness inevitably overlap.

The verdict was unambiguous. The Elastic Net Regressor outperformed its competitors, achieving an R-squared value of 0.9961 and a root mean square error of 0.511. In plain terms, an R-squared approaching 1.0 means the model explains almost all of the variance in the observed academic outcomes, and an RMSE of roughly half a point indicates that its predictions deviate only modestly from actual performance values. Such near-ceiling accuracy on a dataset of this size will invite scrutiny — models that fit training data almost perfectly sometimes reflect overfitting rather than genuine predictive power — but the authors position the result as evidence that the relationships between academic performance and its determinants can be represented effectively by machine learning when the right features are measured and the right constraints are applied. The elastic net’s built-in feature selection, which pushes the coefficients of uninformative variables toward zero, is likely part of the reason it generalized better than the tree-based ensembles on this data.

Perhaps the most consequential aspect of the study is not the raw accuracy but the transparency layer the researchers built on top of it. Black-box predictions have long been a barrier to adopting machine learning in education, because administrators and educators are understandably reluctant to act on a number they cannot interrogate. Madaan and Sharma addressed this by applying two of the most widely used explainable AI frameworks: SHAP, or SHapley Additive exPlanations, which borrows concepts from cooperative game theory to distribute credit for each prediction among the input features, and LIME, or Local Interpretable Model-agnostic Explanations, which builds simple, locally faithful surrogate models around individual predictions. Together, these methods allow the system to show not just that a student is predicted to underperform, but which specific factors — a health issue, a family circumstance, a time-management pattern — are driving that forecast for that particular individual.

This interpretability matters because it converts prediction into action. A model that merely flags a student as at risk tells an advisor where to look; a model that says the risk is driven primarily by, say, inadequate access to campus services or excessive social media consumption tells the advisor what to do. The authors visualize the factors influencing academic performance graphically, and they describe the development of dashboards intended to support decision-making processes within higher education institutions. In operational terms, the envisioned workflow is straightforward: institutions feed student data into the model, the system computes predicted outcomes and their explanations, and staff use those outputs to detect struggling students early, offer timely interventions, and ultimately reduce dropout. The researchers argue that such a system can be beneficial across the sector, from resource-constrained colleges to large research universities.

The study also connects to a lively international research conversation. Recent work has applied machine learning to academic performance prediction across a wide range of contexts — from graph neural networks that model peer influence in online question pools, to exercise-aware knowledge tracing systems that track what a student knows in real time, to comparative studies of classification and regression models among postgraduate public health students. Other teams have explored elastic net approaches in adjacent domains, including using elastic net regression to classify developmental language disorders, underscoring the method’s versatility when datasets contain many correlated predictors. What the Banasthali Vidyapith study adds to this landscape is the combination of a multi-dimensional survey-based feature set, a head-to-head algorithm comparison, and an explicit explainability pipeline, all framed within the sustainability agenda that increasingly shapes higher education policy.

There are, of course, important caveats. A model trained on 200 students at a single institution cannot simply be transplanted to a different university with a different student population, and the authors note that no datasets were generated or analyzed beyond the study itself, with code available on request. Survey-based measures of behavior and psychological traits are self-reported and therefore subject to well-known biases, and any system that predicts student outcomes raises ethical questions about labeling, privacy, and the risk that a low prediction becomes a self-fulfilling prophecy. The researchers are careful to state that academic records were used strictly for research purposes under institutional ethical guidelines, and the growing literature on explainable educational AI reflects a shared awareness that transparency is not just a technical nicety but an ethical requirement when algorithms touch people’s educational trajectories.

Even with those qualifications, the direction of travel is clear. As artificial intelligence matures from novelty to infrastructure in higher education, the institutions that thrive will be those that can turn data into timely, comprehensible, and humane action. A regression model that achieves an R-squared of 0.9961 while explaining its own reasoning through SHAP and LIME offers a glimpse of what that future could look like: not chatbots replacing teachers, but quiet analytical systems working in the background, flagging trouble before it becomes failure, and giving educators the evidence they need to intervene while intervention still matters. For the students who would otherwise have slipped through the cracks, that difference may be the most important number of all.

Subject of Research: Machine learning prediction of university student academic performance using regression models and explainable AI

Article Title: Predicting student academic performance using a regression model

Article References: Madaan, N., & Sharma, S. K. (2026). Predicting student academic performance using a regression model. SN Social Sciences, 6(10), Article 545. https://doi.org/10.1007/s43545-026-01838-2

Image Credits: AI Generated

DOI: 10.1007/s43545-026-01838-2

Keywords: machine learning, educational data mining, academic performance prediction, elastic net regression, explainable AI, SHAP, LIME, higher education, student dropout, predictive analytics, sustainability, regression models

The researchers argue that the level of abstraction at which fault prediction operates matters enormously. Much of the classic literature works at the granularity of individual classes, scoring each class for its likelihood of harboring a defect using metrics such as lines of code, cyclomatic complexity, or the well-known CK metric suite. But modern Java applications are organized into packages, and it is at this architectural level that developers actually plan, review, and refactor their code. Predicting faults at the class level produces thousands of scattered warnings that no team can act on coherently, while predicting at the package level yields a shortlist of architectural units that managers and engineers can directly schedule for inspection, testing, or restructuring. The new model therefore treats the package, not the class, as its unit of analysis, and the authors demonstrate through empirical evaluation on standard benchmark datasets that this package-level abstraction is an effective and practical lens for fault prediction.

What distinguishes this work from earlier package-level studies is its explicit attention to code reuse. Software is rarely written entirely from scratch; developers routinely reuse classes, share functionality across packages, and lean on familiar patterns to accelerate development. Reuse is usually celebrated as a virtue, because well-tested, repeatedly exercised code tends to be stable. But reuse also creates hidden coupling: when several packages depend on the same reused components, a fault in that shared core can propagate widely, and a package whose cohesion depends on reused elements may behave very differently from one built from bespoke code. The authors’ model therefore incorporates architectural metrics designed to capture cohesion, dependency structures, and reuse patterns simultaneously. In doing so it extends a line of research in which Shaikh and colleagues had previously shown that reusability-based package organization carries real signal about where bugs concentrate in object-oriented systems.

The metric suite at the heart of the model draws on decades of work in software modularization research. Cohesion metrics measure how tightly the elements inside a package belong together, on the premise that a package should implement a single, well-defined responsibility. Coupling and dependency metrics count and characterize the relationships between packages, since tangled dependency webs and cyclic references are notorious incubators of defects. Reuse metrics quantify how much of a package’s content is shared with or derived from other parts of the system. The study’s appendices provide precise mathematical definitions of these package-level measures, computed with tooling that includes the Understand static analysis environment, and the authors curated their datasets from widely used open-source Java systems, including Eclipse, Apache POI, Lucene, Camel, and the Eclipse JDT Core, with bug information drawn from established defect repositories used across the fault-prediction community.

On the learning side, the researchers assembled a multi-layered evaluation design rather than relying on a single train-and-test split. Cross-validation is used to check that predictive performance holds up across different partitions of the data rather than reflecting one lucky split. More distinctively, the authors adopt an effort-oriented perspective, following recent trends in just-in-time and effort-aware defect prediction. The insight behind this orientation is that raw accuracy is not the metric that matters to a development team; what matters is how many defects a reviewer can uncover per unit of inspection effort. A model that ranks the most defect-dense packages first delivers more practical value than one with a nominally higher classification score that wastes reviewers’ time on sparsely buggy code. By evaluating their model through this effort-aware lens, the authors align the experimental protocol with the way fault prediction is actually consumed in industrial settings.

The third layer of the approach, and the one with the broadest implications, is explainable AI. The researchers integrate two of the most widely adopted interpretability techniques: SHAP, or SHapley Additive exPlanations, which borrows ideas from cooperative game theory to assign each input feature a fair share of credit or blame for an individual prediction, and LIME, or Local Interpretable Model-agnostic Explanations, which approximates a complex model locally around a given prediction with a simple, human-readable surrogate. Applied to fault prediction, these methods can tell a developer not merely that package X is flagged as fault-prone, but that the flag is driven, for example, by unusually low cohesion combined with heavy dependence on reused components and a dense web of outgoing dependencies. Such explanations transform the model’s output from an opaque verdict into an actionable diagnostic that points toward a specific remedy, whether that is remodularization, targeted testing, or dependency cleanup.

The empirical results, reported across multiple figures and algorithmic descriptions in the paper, support the central conjecture of the study: that package-level abstraction, enriched with reuse awareness and paired with interpretability, improves both the accuracy and the usefulness of fault prediction compared with traditional approaches. The authors note that traditional fault-prediction methods have often lacked scalability, interoperability, and smooth integration into software-quality workflows, and their architecture is designed to address precisely these shortcomings. Just as important, the explanation layer reveals which architectural properties the model has learned to associate with defects, giving the research community a window into the causal texture of software quality rather than a bare leaderboard of predictive scores.

Perhaps the most practically significant finding is the authors’ conjecture that modular design of packages can significantly improve software maintainability and reliability. Because the explainable model surfaces cohesion, dependency, and reuse as the dominant predictors of fault-proneness, it indirectly validates a long-standing principle of software architecture: packages that are cohesive, loosely coupled, and organized around reusable responsibilities tend to be healthier and less buggy. This gives architects empirical, data-driven ammunition when they argue for refactoring a tangled package structure, and it suggests that quality-assurance budgets should be steered toward packages whose architectural metrics deviate from these healthy norms. The work also connects to a growing body of literature on explainable machine learning in software engineering, including studies applying SHAP-style interpretation to just-in-time defect prediction, which together signal a field-wide shift from pure prediction toward prediction that engineers can understand and act upon.

The study was supported by the Sindh Higher Education Commission of Pakistan through its Sindh Research Support Program, by a National Research Foundation of Korea grant funded by the Korean Ministry of Science and ICT, and by an Institute of Information and Communications Technology Planning and Evaluation grant under Korea’s Leading Generative AI Human Resources Development program, reflecting an international collaboration spanning Pakistan, Italy, and South Korea. The datasets and metric definitions described in the manuscript have been made available to the community, including a public repository of the package-level datasets used in the experiments, which lowers the barrier for other groups to replicate and extend the results. For an industry drowning in technical debt, the message of this research is refreshingly concrete: look at how your packages are organized, pay attention to reuse and coupling, let an interpretable model show you where the risk lives, and fix the architecture before the bugs fix your schedule for you.

Subject of Research: Explainable AI-based fault prediction at the package level in object-oriented software

Article Title: Reuse-aware package-level fault prediction with Explainable AI

Article References: Shaikh, M., Akram, S., Tunio, I., Khan, J., & Jung, Y. (2026). Reuse-aware package-level fault prediction with Explainable AI. Cluster Computing, 29(12), Article 705. https://doi.org/10.1007/s10586-026-06509-8

Image Credits: AI Generated

DOI: 10.1007/s10586-026-06509-8

Keywords: software fault prediction, explainable AI, SHAP, LIME, package cohesion, code reuse, software metrics, machine learning, software engineering, Java, maintainability, defect prediction

Soil acidity is one of the most widespread constraints on crop production in tropical highlands, and Ethiopia is no exception. When soils become acidic, aluminum and manganese become soluble and toxic to plant roots, while essential nutrients such as phosphorus are locked into unavailable chemical forms. The molecular basis of this problem is well understood: root growth is inhibited, and the plant’s ability to take up water and nutrients collapses. For wheat, a staple crop that Ethiopia has been striving to produce in sufficient quantities for its growing population, acidic soils translate directly into depressed yields and poor returns on fertilizer investments. Previous work by the same research group has mapped the extent and distribution of surface soil acidity across Ethiopia’s rainfed areas, finding that the problem is extensive and worsening under continuous cultivation.

To untangle the interactions between acidity correction and nutrient supply, the researchers designed a split-plot factorial experiment. Two levels of agricultural lime occupied the main plots, while five nitrogen rates crossed with four phosphorus rates formed the sub-plots. This design allowed the team to measure not only the individual effects of each input but also the interactions among them, across two full growing seasons. They measured soil chemical properties, wheat grain and straw yields, nutrient uptake, nutrient recovery efficiency, and the concentration of zinc in the harvested grain. Zinc was a critical variable because wheat is a major dietary source of this micronutrient in Ethiopia, and zinc deficiency in human populations remains a serious public health concern according to the World Health Organization’s global reporting on micronutrient status.

The headline finding is that nitrogen, not phosphorus or lime, is the dominant yield-limiting nutrient on these acidic soils. Across the experiment, the median yield gain attributable to nitrogen application was 3.5 tonnes per hectare, substantially larger than the 2.0 tonnes per hectare gained from phosphorus and the 1.1 tonnes per hectare gained from liming. In practical terms, this means that a farmer who applies nitrogen fertilizer alone will see the largest single jump in productivity, but the full benefit of that nitrogen can only be captured when soil acidity is corrected and phosphorus supply is adequate. Lime worked largely by improving the soil chemical environment: neutralizing exchangeable acidity, reducing aluminum toxicity, and thereby enhancing the crop’s ability to take up and recover both nitrogen and phosphorus from the soil.

Yields responded strongly at first but showed clear diminishing returns as fertilizer rates climbed. The maximum observed yield in the first season was 4.8 tonnes per hectare, achieved with 150 kilograms of nitrogen and 40 kilograms of phosphorus per hectare. In the second season, the maximum rose to 5.6 tonnes per hectare, but notably this was reached at a lower nitrogen rate of 100 kilograms per hectare with 20 to 40 kilograms of phosphorus. This pattern illustrates a core principle of agronomy: as one constraint is relieved, the efficiency of additional inputs declines, and the economically optimal rate may be well below the agronomic maximum. For resource-poor smallholders, the difference between the last few kilograms of fertilizer and the yield they deliver can determine whether an input package pays for itself at all.

The most provocative result, however, concerns grain zinc concentration. The researchers uncovered a trade-off between wheat yield and grain nutritional quality that was independent of lime and nitrogen management: as phosphorus application increased, yields rose but grain zinc concentrations fell. Across the experiment, grain zinc rarely exceeded 35 milligrams per kilogram, a threshold relevant to biofortification targets for human nutrition. Lime, by contrast, increased grain zinc concentration, while excessive nitrogen and incremental increases in phosphorus rates both reduced it. This dilution effect, in which larger grain harvests spread a finite pool of micronutrients across more biomass, is a recognized phenomenon in crop physiology, but seeing it play out so clearly against a background of soil acidity correction is a warning for breeding and fertilizer programs alike.

The implications for policy are twofold. First, the study supports the integration of lime with balanced fertilization as the backbone of sustainable intensification in the Ethiopian highlands. Liming is expensive and logistically challenging, and recent analyses in Nature Food have argued that soil acidity remediation across sub-Saharan Africa requires targeted rather than blanket investments. This experiment provides the field-level evidence for such targeting: lime pays off most where it unlocks the response to nitrogen and phosphorus, and its benefits extend beyond yield to include improved nutrient recovery efficiency, meaning less fertilizer is wasted and less is lost to the environment. Second, the yield-versus-zinc trade-off means that cautious phosphorus management is needed, and complementary zinc supplementation, whether through foliar zinc sprays, zinc-enriched fertilizers, or zinc-biofortified wheat varieties, will be required to keep grain quality on track.

The research also speaks to a broader debate about how Africa achieves wheat self-sufficiency. Ethiopia’s national strategy has pursued yield gap closure, land expansion and irrigation, but the yield gap studies led by co-author João Vasco Silva have shown that soil fertility constraints, including acidity, are among the chief culprits keeping actual yields far below potential. By quantifying the marginal productivity of each input and their interactions, this experiment gives Ethiopian extension services and fertilizer blending programs an empirical basis for recommending site-specific packages. The finding that the second year’s maximum yield came at lower nitrogen rates than the first also hints at residual effects of lime and possibly carryover phosphorus, a reminder that soil fertility management is a multi-year enterprise rather than a season-by-season calculation.

Methodologically, the study is notable for its use of nutrient recovery efficiency as a diagnostic, following frameworks established by Dobermann and by the QUEFTS model for quantitative evaluation of tropical soil fertility. Recovery efficiency measures the fraction of applied nutrient that actually ends up in the aboveground crop, and it is a sensitive indicator of whether the soil environment is allowing roots to function. The observation that lime enhanced both uptake and recovery of nitrogen and phosphorus confirms the mechanistic chain: acidity correction improves root health, which improves capture of applied nutrients, which improves yield. The statistical analysis relied on linear mixed-effects models and quantile regression, tools that let the researchers characterize responses across the full distribution of plot-level outcomes rather than only at the mean.

For the farmers of the Upper Blue Nile Basin, the practical message is a recipe: correct the acidity first, supply nitrogen as the primary driver of yield, apply phosphorus judiciously rather than maximally, and plan for zinc. For scientists and breeders, the message is that productivity and nutrition goals must be evaluated together, because a fertilizer recommendation that maximizes tonnes per hectare may quietly erode the micronutrient content of the food that same hectare produces. As Ethiopia and its neighbors invest heavily in soil health and fertilizer value chains, this study demonstrates that the cheapest route to both more food and better food runs through the chemistry of the soil itself, and that ignoring the trade-offs hidden in that chemistry would be a costly mistake.

Subject of Research: Effects of lime, nitrogen and phosphorus management on wheat yield, nutrient use efficiency and grain zinc concentration in acidic soils of the Upper Blue Nile Basin, Ethiopia

Article Title: Nutrient management strategies to balance wheat productivity, grain quality and fertilizer use efficiency on acidic soils in the Upper Blue Nile Basin, Ethiopia

Article References: Lulie, B., Adgo, E., Desalegn, T., Sida, T. S., & Silva, J. V. (2026). Nutrient management strategies to balance wheat productivity, grain quality and fertilizer use efficiency on acidic soils in the Upper Blue Nile Basin, Ethiopia. Plant and Soil. https://doi.org/10.1007/s11104-026-09064-8

Image Credits: AI Generated

DOI: 10.1007/s11104-026-09064-8

Keywords: wheat, soil acidity, liming, nitrogen, phosphorus, zinc, nutrient use efficiency, Ethiopia, Upper Blue Nile Basin, grain quality, sustainable intensification, soil fertility

The central question the researchers posed is deceptively simple: when every weather sensor on a station is reporting, which of those readings genuinely carry information about incoming solar radiation, and which are noise, redundancy, or artifacts of imperfect instrumentation? Conventional practice often throws every available variable into a machine learning model and lets the algorithm sort it out. That approach can work, but it carries a cost. Models trained on irrelevant or collinear predictors tend to be fragile, harder to interpret, and more likely to break down when a sensor malfunctions or conditions shift outside the training range. In renewable energy applications, where forecasts feed directly into grid scheduling decisions, that fragility is not an academic concern—it is an operational liability.

To address the challenge, the team built a multi-stage evaluation pipeline that deliberately combines three complementary families of analysis. The first stage probes linear structure in the data. The researchers applied three rank- and product-moment correlation methods—Pearson, Spearman, and Kendall—to quantify how tightly each meteorological predictor co-varies with solar irradiance, and they deployed principal component analysis (PCA) to map the dominant axes of variance running through the whole dataset. PCA is a classical dimensionality reduction technique that rotates the data into new coordinates ordered by the amount of variance each captures; when a handful of components account for most of the spread, it signals that many original variables are telling overlapping stories. In this study, the first three principal components together explained 62.88 percent of the total variance, a clear indication of substantial multicollinearity among the sensor readings.

Linear analysis alone, however, cannot capture the full picture. Atmospheric variables interact in ways that are stubbornly nonlinear: humidity modulates cloud formation, which alters irradiance, but the relationship bends and saturates depending on temperature, season, and pressure. To test nonlinear sensitivity, the team trained a bilayer neural network—a shallow architecture with two hidden processing layers—capable of registering interactions that straight-line correlation would miss. Crucially, the neural network served not as a final forecasting product but as a probe: by measuring how prediction error responds when each input is perturbed, the researchers could rank variables by their functional importance rather than their statistical association alone. This dual-track strategy, pairing classical statistics with neural sensitivity testing, forms the hybrid core of the framework.

The third and arguably most novel stage brought explainable artificial intelligence into the mix. The researchers applied SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations), two techniques borrowed from the machine learning interpretability literature that decompose a model’s prediction into contributions from each input variable. SHAP, grounded in cooperative game theory, allocates to each feature a Shapley value representing its average marginal contribution across all possible coalitions of predictors, yielding a global picture of which variables drive the model overall. LIME, by contrast, fits simple surrogate models around individual predictions to explain why a specific forecast took the value it did. Combining the two gives both the bird’s-eye view and the close-up, allowing the team to verify that the variables flagged by correlation and neural sensitivity are the same ones the trained model actually leans on when making its estimates.

The results delivered a clear hierarchy of predictors. The most influential single variable turned out to be average battery voltage, abbreviated ABV, which achieved a root mean square error of 0.203 when used to predict solar irradiance. The finding may look surprising at first glance—an electrical quantity appearing to outperform classic atmospheric measurements—but it makes physical sense in context: at a ground-based monitoring station, battery voltage tracks the charging cycle driven by the panel’s own exposure to sunlight, making it a sensitive, integrated proxy for the very quantity being forecast. Relative humidity followed as the second-strongest predictor, with an RMSE of 0.239, consistent with its well-documented role in cloud and aerosol processes that scatter and absorb incoming radiation. Wind direction ranked third at 0.244, plausibly reflecting its association with air mass origin and the transport of moisture and particulates.

The explainability analysis reinforced this ranking from an independent angle. Global SHAP assessment confirmed that ABV and relative humidity contributed most significantly to the model’s predictions, while rainfall emerged as a weak global influence—a result the researchers attribute to its sparse occurrence in the record. Rainfall events, though meteorologically dramatic, occupy so few rows in a typical time series that the model has little opportunity to learn a robust mapping from rainfall to irradiance, and the feature’s Shapley values remain correspondingly small. This kind of diagnostic insight is precisely what the framework is designed to surface: rather than silently retaining a predictor that adds little, reliability-aware selection recognizes that a rarely active sensor contributes maintenance cost and potential failure modes without commensurate predictive benefit.

Beyond the rankings themselves, the study’s methodological contribution lies in its explicit handling of the data-quality hazards that plague real-world meteorological records. The authors frame the problem as one of identifying physically meaningful predictors under conditions that are simultaneously non-normal, multicollinear, and sensor-sensitive—three threats that conventional single-method feature selection handles poorly. Non-normal distributions undermine parametric tests; multicollinearity inflates the apparent importance of redundant variables and destabilizes coefficient estimates; and sensor sensitivity means that measured values may reflect instrument behavior as much as atmospheric reality. By triangulating across correlation families, PCA, neural sensitivity, and interpretable machine learning, the framework forces a candidate predictor to earn its place through convergent evidence before it is accepted into a forecasting model.

The practical implications extend across the renewable energy value chain. More reliable irradiance forecasts translate directly into better solar farm siting, more accurate energy yield assessments for investors, and smoother integration of photovoltaic capacity into smart grids, where dispatch decisions must anticipate the variable output of solar assets. The framework’s emphasis on lean, interpretable predictor sets also lowers computational and maintenance burdens, since models built on a handful of well-justified inputs are cheaper to run and easier to audit than sprawling ones. For operators in data-scarce regions, the finding that a simple, routinely logged variable such as battery voltage carries exceptional predictive power suggests that meaningful forecasting capability may be achievable with minimal instrumentation—a potentially significant result for expanding solar infrastructure across developing regions where dense sensor networks are not yet in place.

The study, which received funding support from a Universiti Sains Malaysia Research University Grant, arrives at a moment when the global build-out of solar capacity is accelerating and the marginal value of forecast accuracy is rising accordingly. It also joins a broader movement in environmental machine learning toward transparency: rather than treating predictive models as black boxes, the authors demonstrate how explanation tools can be woven into the model-building process itself, serving as a quality check on the science as much as a window on the model. For a field where forecasting errors carry real economic and grid-stability consequences, the message is clear—the path to trustworthy solar prediction runs not through bigger models with more inputs, but through disciplined, reliability-aware scrutiny of the measurements that feed them.

Subject of Research: Hybrid feature selection for ground-based solar irradiance prediction using meteorological data and explainable machine learning

Article Title: Hybrid feature selection framework for reliable solar irradiance prediction using ground-based meteorological data

Article References: Maiwada, A. D., Ahmad, M. I., Abba, S. I., & Shakir, M. A. (2026). Hybrid feature selection framework for reliable solar irradiance prediction using ground-based meteorological data. Discover Sustainability. https://doi.org/10.1007/s43621-026-04977-x

Image Credits: AI Generated

DOI: 10.1007/s43621-026-04977-x

Keywords: solar irradiance, feature selection, machine learning, SHAP, LIME, PCA, neural networks, renewable energy, meteorology, forecasting, multicollinearity, smart grids