For the analysis, the authors harmonized Landsat 8/9 land surface temperature, Dynamic World land-cover probabilities, JRC surface water, WorldPop population, and Global Human Settlement Layer morphology on a 500 m equal-area grid. They combined residentially weighted indicators, sensitivity tests, an exact-count 3 km spatial block null, and spatially validated gradient boosting. ERA5-Land was used only for city-level summaries of broad climate context and was excluded from grid-level prediction. The abstract reports that the highest population-density decile was 4.19–5.23 °C warmer than the lowest decile, with the same direction observed under alternative population and boundary specifications. Gradients involving older adults varied among the four cities.

Primary spatial cross-validation produced coefficients of determination of 0.632–0.896 for the city models and 0.880 for the pooled model. Built form, Landsat spectral variables, and pooled city indicators were the dominant predictors. Blue-green variables made smaller conditional contributions to prediction, and transfer to fully held-out years or cities was limited. These results describe predictive associations within the analyzed grids; they do not establish that population density or blue-green cover causes the observed temperature differences.

Because the supplied material consists of the abstract and a subscription preview rather than the full article, the reported methods, numerical results, and limitations above are limited to information stated in the abstract. The authors describe the framework as a repeatable way to identify persistent residential surface-heat and blue-green-resource mismatches for targeted follow-up, while the limited temporal and cross-city transfer constrains how broadly its predictions can be applied.

Subject of Research: Satellite-based assessment of residential surface heat and blue-green resource mismatch in four Chinese cities

Article Title: Satellite monitoring of population-weighted surface heat and blue-green resource mismatch across four Chinese cities

Article References: Pu, W., Li, G., Tan, X., & Li, Z. (2026). Satellite monitoring of population-weighted surface heat and blue-green resource mismatch across four Chinese cities. Environmental Monitoring and Assessment, 198(11), Article 1212. https://doi.org/10.1007/s10661-026-16084-5

Image Credits: AI Generated

DOI: 10.1007/s10661-026-16084-5

Keywords: urban heat, satellite monitoring, blue-green space, population density, land surface temperature, spatial modeling, gradient boosting, Harbin, Beijing, Wuhan, Shenzhen, Satellite

Rather than merging these feature families into a single vector space, the model trains each branch separately, converts their outputs into comparable class probabilities, and blends them with a weighted average whose mixing coefficient is tuned on a validation set. When the two branches were given equal weight, performance actually dropped; only when the much stronger text branch dominated did fusion deliver a benefit. On a held-out test set, HTSELM achieved an accuracy of 0.9392, a macro-F1 score of 0.9398, a Matthews correlation coefficient of 0.9043 and a macro ROC-AUC of 0.9923.

Exact five-fold cross-validation produced a mean macro-F1 of 0.9398 with a standard deviation of just 0.0069, indicating that the result was stable rather than a lucky split. The improvement over a text-only support vector machine was small—0.0048 in macro-F1—but consistent across all five cross-validation folds. The most provocative part of the study, however, is not the headline numbers but the audit that preceded them.

Wang examined the Social Media Engagement Dataset, a machine-generated collection of 12,000 records with 28 raw columns, and discovered that the released sentiment_score variable deterministically defines the sentiment_label target: negative labels correspond to scores below roughly −0.20, neutral labels fall in a narrow band around zero, and positive labels begin above roughly +0.20. Similarly, the engagement_rate column agreed perfectly, within rounding tolerance, with a formula computed from likes, shares, comments and impressions. Any model trained with these variables included would not be predicting anything—it would simply be reconstructing the answer key.

By removing sentiment_score, emotion_type, user_past_sentiment_avg and toxicity_score before training the sentiment classifier, and by excluding engagement_rate and all post-response variables from the engagement experiment, the study establishes what its author calls a leakage-controlled baseline—an honest floor rather than an inflated ceiling. The ablation experiments reinforce how much of the predictive signal lives in the words themselves. The explainability analysis tells a consistent story: the most influential TF-IDF terms included phrases such as "okay," "wouldn't recommend" and "highly recommend," recognisable polarity expressions that dominate the classification decision, whereas permutation importance for the structured features was small and dispersed across brand, product, language, campaign and location variables.

In other words, once affective proxies are stripped away, metadata contributes surprisingly little to sentiment discrimination in this corpus. Robustness testing added further nuance. Perturbing numeric features with noise barely dented performance, with macro-F1 values ranging only from 0.9390 to 0.9423 across perturbation levels.

Missing features and text corruption, by contrast, caused gradual but meaningful degradation: at 20 percent perturbation, missing-feature tests dropped macro-F1 to 0.8441 and token-level text noise reduced it to 0.8599. The practical implication is clear—operational reliability of such systems depends far more on the availability and integrity of the text itself than on the precision of the accompanying numbers. The confusion matrix showed the model handled the Neutral class reasonably well, with a recall of 0.9069, and that residual errors were mostly polarity confusions between negative and positive posts.

Perhaps the study's most sobering finding comes from its secondary experiment. When the framework attempted to predict engagement categories using only information available before an audience responds—text, platform, timing and campaign metadata, with all response-derived counters excluded—every model performed close to the three-class chance level of one-third. The best configuration, a structured logistic regression, managed an accuracy of just 0.3479, and HTSELM itself reached only 0.3413 with a Matthews correlation coefficient of 0.0118, effectively zero.

Wang argues this negative result is scientifically valuable: it demonstrates that the pre-response variables in this dataset carry no reliable predictive signal, and it prevents future work from mistaking formula recovery for genuine forecasting ability. The author is careful about the limits of these conclusions. The dataset is synthetic and machine-generated, so its linguistic templates, platform distributions and behavioral patterns may not mirror naturally occurring social media.

The baseline set deliberately excludes transformer models such as BERT and RoBERTa, which could offer richer contextual representations of sarcasm, negation and code-mixing, though the study notes that any such comparison should first be subjected to the same leakage controls. Statistical tests also revealed the constraints of a small evaluation: paired t-tests across five folds favored HTSELM over the text-only baselines, but Wilcoxon and McNemar tests did not reach significance, reflecting limited statistical power rather than contradictory evidence. The study frames explainability not as an afterthought but as part of the experimental protocol, reporting coefficient analyses, permutation importances and error shares alongside performance metrics, in line with human-centered AI principles that demand transparency about data provenance, model limitations and deployment boundaries.

Subject of Research: Machine learning analysis of social media sentiment and engagement with leakage control

Article Title: Artificial intelligence-based analysis of digital media engagement and user sentiment using machine learning techniques

Article References: Wang, X. (2026). Artificial intelligence-based analysis of digital media engagement and user sentiment using machine learning techniques. Discover Artificial Intelligence, 6(1), Article 1448. https://doi.org/10.1007/s44163-026-02430-6

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02430-6

Keywords: artificial intelligence, machine learning, sentiment analysis, social media, target leakage, engagement prediction, late fusion, explainable AI, TF-IDF, gradient boosting, digital media analytics, data leakage

Patients in the study had both lower extremity arterial disease, involving stenosis or occlusion of arteries supplying the legs, and diabetes. Their mean age was 72.7 years, and 65.7% were men. The investigators divided the cohort into 967 training participants, 414 internal-validation participants and 538 external-validation participants.

During the standardized 12-month observation period, 269 patients underwent major or minor amputation, 118 died before amputation without limb loss, and 1,532 were censored. Amputation occurred in 14.6%, 13.3% and 13.6% of the training, internal-validation and external-validation groups, respectively. A competing event matters because a conventional binary model may treat death before amputation as simply a non-amputation outcome, potentially distorting the estimated probability that a patient actually undergoes limb loss while remaining at risk.

To reduce information leakage, candidate predictors were screened using only the training cohort. A systematic literature review initially identified 75 potential amputation-related factors, after which clinical accessibility and data completeness reduced the set to 48 variables. Three complementary methods—LASSO logistic regression, a random forest classifier and the Boruta algorithm—then ranked or selected those candidates.

A variable was retained when at least two methods supported it, yielding 14 predictors: Rutherford classification, financial burden, age, Barthel Index, diabetic foot, statin use, fibrinogen, neutrophil count, residence, prealbumin, C-reactive protein, activated partial thromboplastin time, body-mass-index classification and the presence of a wound. The same predictor set was used for both the machine-learning and competing-risk models, allowing their performance to be compared under a common population and 12-month prediction window. Ten algorithms were then trained with hyperparameter tuning and evaluated on discrimination, calibration and related classification measures.

The candidates were random forest, logistic regression, decision tree, naive Bayes, artificial neural network, k-nearest neighbors, XGBoost, gradient boosting, CatBoost and light gradient boosting machine. The selected GraBoost model, a gradient-boosting algorithm, performed best in internal validation: its AUC was 0.860, accuracy was 0.787, F1 score was 0.782 and Brier score was 0.156. In the geographically separate external cohort, it retained an AUC of 0.828, accuracy of 0.760, F1 score of 0.695 and Brier score of 0.178.

Eight algorithms had external AUC values above 0.70, while decision tree and k-nearest neighbors performed less consistently. DeLong tests found statistically significant AUC differences between GraBoost and most competing models in internal validation, although some external differences were not significant. These results describe discrimination and calibration in the studied cohorts; they do not establish that the model can improve clinical decisions or patient outcomes.

SHAP analysis, which decomposes each prediction into the directional contribution of its inputs, showed that Rutherford classification had the strongest influence, followed by financial burden and age. Barthel Index, diabetic foot, statin use, fibrinogen, neutrophil count, residence and prealbumin had moderate effects, while C-reactive protein, activated partial thromboplastin time, body-mass-index classification and wounds had comparatively weaker associations. Higher Rutherford grades, lower Barthel Index scores, younger age, wounds and diabetic foot were associated with higher predicted amputation risk.

Higher fibrinogen, activated partial thromboplastin time, C-reactive protein and neutrophil values also contributed to higher predictions, whereas higher prealbumin was associated with lower risk. Urban residence, lower financial burden, normal body-mass-index classification and statin use contributed in the protective direction. These are predictive associations in an observational cohort, not evidence that changing any one factor would cause a reduction in amputation.

The Fine-Gray competing-risk regression used the same 14 inputs and treated amputation as the index event and all-cause death before amputation as the competing event. Its 12-month time-dependent AUC was 0.762 in the training cohort, 0.711 in internal validation and 0.811 in external validation. Corresponding Brier scores were 0.092, 0.106 and 0.085.

Calibration curves for 3-, 6- and 12-month endpoints indicated agreement between predicted and observed cumulative incidence across risk strata, while Gray’s test found no significant difference in 12-month cumulative amputation incidence among the three cohorts, with p = 0.722. The relatively small internal-validation group produced wide confidence intervals for the time-dependent AUC, adding uncertainty to the estimate of discrimination. The authors also generated a nomogram that assigns points to the 14 predictors and translates total points into estimated 3-, 6- and 12-month cumulative incidence under the competing-risk framework.

Several patterns challenge simple assumptions about risk. Younger age was associated with greater predicted amputation despite the general tendency for arterial disease to become more prevalent with age. The investigators proposed possible explanations, including more aggressive surgical decisions among younger patients with severe limb complications and a more aggressive disease phenotype in some younger patients, but emphasized that these ideas are speculative.

They also noted that residual confounding could not be excluded. Similarly, the observed association between rural or town residence, greater financial burden and amputation may reflect differences in access to care, delayed intervention or adherence, although the retrospective data cannot establish those mechanisms. The model also identified a U-shaped pattern in which both underweight and higher body-mass-index classifications were associated with greater predicted risk.

Underweight may reflect malnutrition, sarcopenia or underlying disease burden, while higher body mass index may accompany metabolic disturbance and inflammation, but the study did not test those pathways as causes of amputation. Wounds and diabetic foot were positively associated with predicted risk, but their SHAP effects were weaker than those of Rutherford classification, financial burden and age. Their presence can be consistent with tissue ischemia, impaired healing and infection, but the model’s outputs do not determine why an individual patient developed either condition.

Statin use was associated with lower predicted risk, while higher fibrinogen and prolonged activated partial thromboplastin time were associated with higher risk. The authors interpreted these laboratory patterns in relation to inflammation, coagulation and tissue healing. They also emphasized the role of limited self-care reflected by a lower Barthel Index: reduced mobility and poorer self-management might delay recognition or treatment of foot injury, although the study did not directly test that pathway.

These biological and social interpretations remain explanations for model associations rather than demonstrated causal effects. The researchers deployed both models in a Shiny-based web calculator that accepts patient information and displays synchronized 12-month risk estimates from the binary GraBoost and Fine-Gray frameworks. Dropdown menus are provided for categorical predictors and sliders for continuous variables, allowing a user to enter clinical data and calculate the two outputs.

The 14 inputs were described by the authors as routinely measurable across different hospital tiers, supporting the practical feasibility of collecting the required data. The paired outputs are intended to present two perspectives: a binary machine-learning estimate and a competing-risk estimate that explicitly accounts for pre-amputation death. Nevertheless, the web implementation has not undergone prospective validation in routine care.

It cannot replace comprehensive clinical assessment, and the study did not test whether using it changes surveillance, revascularization, medication prescribing, limb outcomes or survival. Important limitations constrain how the findings should be interpreted. The study was retrospective and observational, so selection bias, incomplete records and unmeasured confounding remain possible.

Baseline characteristics differed among the three cohorts, including differences in Rutherford classification, Barthel Index, residence, statin use and other variables. All participants were Chinese patients treated in Chinese hospitals, limiting extrapolation to other ethnic populations and healthcare systems. Multiple imputation was performed on the pooled dataset before partitioning, which may have introduced some data leakage even though feature selection was restricted to the training cohort.

Patients lost to follow-up also differed significantly from the overall cohort on most baseline variables, indicating non-random attrition. The models were designed specifically for 12-month amputation and cannot be assumed to predict longer-term outcomes. The observed 14.0% amputation rate was also within the 13%–48% range reported for comparable populations, although variations among studies may reflect differences in inclusion criteria, disease severity, healthcare systems and follow-up definitions.

Taken together, the study demonstrates promising external discrimination in a defined clinical population, while prospective multicenter testing remains necessary before the tools can support routine clinical use.

Subject of Research: Development and external validation of 12-month amputation risk models for patients with lower extremity arterial disease and diabetes

Article Title: Development and validation of a machine learning-based model for predicting amputation risk in patients with lower extremity arterial disease and diabetes

Article References: Tang, L., Wang, L., Li, J., Liu, X., Chen, Q., Song, P., Quan, X., Li, Q., Liu, G., Liu, Y., Cui, C., & Shi, M. (2026). Development and validation of a machine learning-based model for predicting amputation risk in patients with lower extremity arterial disease and diabetes. Frontiers in Cardiovascular Medicine, 13, Article 1893875. https://doi.org/10.3389/fcvm.2026.1893875

Image Credits: AI Generated

DOI: 10.3389/fcvm.2026.1893875

Keywords: amputation risk, lower extremity arterial disease, diabetes, machine learning, gradient boosting, competing risks, external validation, clinical prediction, SHAP interpretability, web calculator, Development, validation

Liposomes are lipid vesicles used to carry medicines and other therapeutic payloads. Their size, morphology, composition and stability can affect how they interact with cells, circulate in the bloodstream, release their cargo and withstand environmental stresses. These properties are therefore treated as critical quality attributes in the manufacture of liposomal drugs. Microfluidic systems, which guide liquids through tiny channels, have become an attractive production platform because researchers can adjust total flow rate, the ratio between the two liquid streams, solvent conditions and channel geometry to influence lipid self-assembly in a controlled environment.

Not all liposomes are the same size, and that variation matters. PDI is typically derived from dynamic light-scattering measurements and describes the breadth of a particle-size distribution rather than its average. A low PDI indicates a comparatively uniform population, whereas a higher value indicates a broader range of vesicle sizes. Broad distributions can make it harder to correlate dose with biological response and can contribute to problems with encapsulation, payload release, stability and sensitivity to temperature changes, residual solvent or purification. The variables that determine liposome dispersion are interconnected: fluid dynamics, diffusion-limited solvent exchange, lipid phase behavior and downstream processing can act together, making conventional one-factor-at-a-time experiments incomplete.

The researchers addressed that challenge by building a data set linking formulation and operating variables to two measurable outcomes: particle size, expressed in nanometers, and PDI. The original inputs included total flow rate, the flow-rate ratio between the two liquid streams, the concentrations of egg sphingomyelin, hydrogenated soy phosphatidylcholine, cholesterol and PEG-lipid, the identity of the aqueous phase, either ultrapure water or phosphate-buffered saline, and the type of microfluidic chip. The data were divided into 214 training samples, 46 validation samples and 46 test samples. Twelve additional wet-laboratory samples were reserved as an external validation set, although the reported results do not state how the models performed on that set.

To give the algorithms more useful information, the team expanded the original nine predictors to 19 engineered variables. These included the fractions of cholesterol, PEG-lipid, sphingomyelin and the principal membrane-forming lipid; the absolute concentration of the main lipid; interaction terms such as cholesterol multiplied by PEG and total flow rate multiplied by flow-rate ratio; and squared terms intended to capture nonlinear effects. The transformations were calculated directly from the measured inputs, allowing the resulting relationships to remain linked to recognizable aspects of liposome self-assembly rather than turning the analysis into an opaque numerical exercise.

Eleven predictive algorithms were evaluated, including ridge, lasso and elastic-net regression; radial-basis support vector regression; random forest, gradient boosting, XGBoost, LightGBM and CatBoost; and shallow and deep multilayer perceptrons. Hyperparameters were selected through grid search with fivefold cross-validation, while neural networks used the Adam optimizer and early stopping based on validation loss. Performance was assessed with the coefficient of determination, or R², root mean squared error and mean absolute error. The models were also examined with permutation importance and SHAP, a model-agnostic method that decomposes individual predictions into the contributions of different features and can provide global feature rankings as well as local explanations.

PDI proved to be the more predictable target. Gradient boosting achieved a validation R² of 0.7650, an RMSE of 0.1058 and an MAE of 0.0589. Its predictions followed the ideal diagonal across PDI values from 0.05 to 0.80 with little systematic bias. Random forest ranked second for PDI with a validation R² of 0.6653, followed by XGBoost at 0.6192. The leading gradient-boosting result also exceeded one previously reported mixture-of-experts benchmark with an R² of 0.7307. This performance is consistent with the ability of tree ensembles to capture nonlinear relationships and interactions in modest-size tabular data without requiring the very large samples often needed by deep neural networks. XGBoost’s very high reported training fit, with a training R² of 0.99, also suggested that it could be at risk of overfitting given the limited data set.

By contrast, the best model for particle size was a deep multilayer perceptron, but its validation R² reached only 0.4598, with an RMSE of 368.42 nanometers and an MAE of 272.83 nanometers. The size data were highly skewed, with a reported skewness of 5.70, and 9.2% of samples were classified as outliers. Predictions became especially variable above 500 nanometers. The authors noted that size can change sharply when formulations move between operating regimes, including when they cross a lipid-concentration threshold, undergo a phase transition or experience changes in solvent-composition kinetics. The size distribution contained a few large-size formulations and a longer tail of smaller-size formulations, a pattern that is difficult to capture reliably with a limited amount of data.

The interpretability analysis nevertheless produced useful clues. The primary lipid proportion ranked first in feature importance, followed by the PEG-lipid ratio, sphingomyelin fraction, flow-rate ratio and cholesterol fraction. SHAP analysis indicated that higher values of the main-lipid ratio were generally associated with lower PDI. PEG-lipid was also important, but at high concentrations its relationship with uniformity became detrimental. The results indicate that composition ratios and flow dynamics are not interchangeable controls: developers may need to adjust the balance among membrane constituents while simultaneously setting the rate at which the streams mix to achieve a narrow distribution. These are model-derived associations, not evidence that any single factor has been proven to cause a particular biological or manufacturing outcome.

The immediate application is model-assisted screening rather than fully autonomous formulation. The strongest PDI predictor could first identify regions of the formulation space that merit laboratory attention, after which targeted experiments could refine both size and dispersion. The authors also point to logarithmic size targets, robust loss functions, quantile or interval predictions, multi-task learning and explicit uncertainty estimates as possible ways to improve the weaker size model. More data are especially needed for large-particle and underrepresented operating regimes, while reporting performance on the 12-sample external validation would help establish whether predictions transfer across batches, operators, chips and measurement conditions. The study therefore does not remove the need for experimental testing, but it offers a way to make experiments more informed and to identify candidate conditions for narrower liposome populations without yet establishing improved clinical performance or product outcomes.

Subject of Research: Explainable machine learning for predicting liposome polydispersity in microfluidic manufacturing

Article Title: Data-driven prediction of polydispersity in microfluidic liposome production for pharmaceutical formulation optimization

Article References: Olamat, A., Abu Owida, H., Arabiat, A., & Ozel, P. (2026). Data-driven prediction of polydispersity in microfluidic liposome production for pharmaceutical formulation optimization. Pharmacia, 73, Article e200497. https://doi.org/10.3897/pharmacia.73.e200497

Image Credits: AI Generated

DOI: 10.3897/pharmacia.73.e200497

Keywords: liposomes, microfluidics, machine learning, polydispersity index, pharmaceutical formulation, gradient boosting, particle size, process optimization, SHAP analysis, drug delivery, Data-driven, prediction

The motivation is straightforward. In developing countries, expensive disease-specific tests are often out of reach for large portions of the population, while the CBC test, ranging from roughly 600 to 1000 Pakistani rupees, is widely affordable. Its twelve standard parameters encode a surprising amount of clinical signal: white blood cell counts point to infection and leukemia, hemoglobin, red blood cell count and hematocrit indicate anemia and polycythemia, platelet counts reveal bleeding and inflammatory disorders, and the differential leukocyte fractions help identify bacterial infection, bone marrow suppression and allergic responses. Yet in most hospitals these reports exist only as paper printouts, unread by any algorithm and difficult to store or analyze at scale.

To build their dataset, the researchers collected nearly 8,000 real CBC reports from 2023 and 2024 at a local hospital, all in PDF form with identical layouts. After applying inclusion and exclusion criteria, 6,500 patient records remained. Crucially, all patient demographics and personal identifiers were removed in compliance with privacy guidelines, and ethical clearance was obtained from the hospital. The team then benchmarked six OCR engines on the scanned pages, including PyMuPDF, OpenCV, PdfMiner, Tesseract, EasyOCR and PaddleOCR, measuring each with the character error rate, the standard metric that counts substitutions, insertions and deletions against a manually verified ground truth.

PaddleOCR emerged as the clear winner, achieving 99 percent accuracy with a character error rate of just 0.01, compared to 93 percent for EasyOCR and 84 percent for PyMuPDF and OpenCV. The winning pipeline involved converting PDFs to images, then applying Gaussian filtering to remove high-frequency noise, median filtering to suppress salt-and-pepper artifacts without blurring character edges, skew correction to align table boundaries, histogram equalization to enhance contrast, and min-max normalization to rescale pixel intensities. A detection model localized the table region and identified rows, columns and value positions, while a CRNN-based recognition model with a structure-preserving module ensured that each extracted value stayed aligned with its correct test name, preventing records from being mixed together.

With the table data digitized, the team faced a labeling problem: the extracted records had no disease annotations. They solved this with a rule-based algorithm built on clinically accepted reference ranges verified by medical experts at the source hospital. Each of the twelve parameters received a dictionary of thresholds with disease associations for values above and below normal. A hemoglobin below 13 g/dL, for example, flags anemia, while a value above 18 g/dL suggests polycythemia; an MCV under 75 fL indicates microcytic anemia, an MCHC above 35 g/dL indicates hyperchromic anemia, platelets below 150 per microliter signal thrombocytopenia, and elevated neutrophils point to acute bacterial infection. The algorithm processed every record against these rules, generating both a binary disease-or-normal label and, importantly, multiple simultaneous disease labels per patient.

The data was then cleaned through exploratory analysis, with outliers detected in the platelet feature capped using a Z-score threshold to preserve clinical relevance, and all features standardized before modeling. For the multi-disease task, the team used multi-label classification, decomposing the single multi-disease output column into seventeen separate binary target columns via a Multi-Label Binarizer, so that each record became a vector of yes-or-no flags for each condition. Five classifiers were trained: logistic regression, decision tree, random forest, gradient boosting and support vector machine, all implemented in scikit-learn on a 16 GB Windows machine running Python 3.12.2.

Hyperparameter tuning proved decisive. Using GridSearchCV integrated with 10-fold cross-validation on an 80:20 train-test split, the team searched parameter grids for each model, setting the decision tree maximum depth to 20 with the Gini criterion, constraining the random forest to a depth of three, six leaf nodes and square-root feature sampling to prevent overfitting, fixing gradient boosting at 100 estimators with a learning rate of 0.1, and configuring the support vector classifier with a regularization parameter C of 3, scaled gamma and a radial basis function kernel. Untuned models peaked at around 74 percent F1-score for binary classification, but after tuning the cross-validated results climbed substantially, with gradient boosting reaching an F1 of 89.4 percent and an ROC-AUC of 0.94.

On the held-out test set, random forest delivered the best binary performance with 91.7 percent precision, 90.9 percent recall and a 0.94 ROC-AUC, narrowly edging out gradient boosting. The multi-label results told a similar story: random forest achieved around 93 percent precision and an F1 of 92.0 percent on cross-validation, with gradient boosting close behind, while logistic regression and the decision tree lagged. The authors noted that misclassified cases tended to involve borderline values only slightly outside normal ranges, highlighting how sensitive threshold-based labeling can be when clinical signals overlap. Across both tasks, the ensemble methods that aggregate many decision trees proved most robust to this ambiguity.

Transparency was addressed with explainable AI. The team applied SHAP, or SHapley Additive Explanations, analysis independently to each disease label, ranking features by their marginal contribution to each prediction. The analysis identified white blood cells, red blood cells, hemoglobin, MCV, platelets, hematocrit, MCH, MCHC and neutrophils as the most influential parameters driving both binary and multi-disease classification. This patient-level interpretability, the researchers argue, is essential for clinician trust in high-stakes healthcare settings, allowing a doctor to see exactly which blood values pushed a given prediction.

The study is the first of its kind, the authors say, to tackle multi-disease prediction from CBC features using a large real-world dataset, and it arrives with clear caveats. The data comes from a single hospital region and only twelve parameters, so results cannot yet be generalized to broader populations, and rule-based labeling may miss complex diagnostic patterns that only appear in richer clinical data. The team is already working to integrate the system into an Android application so patients and doctors can interpret reports interactively, and they suggest that hospitals should maintain online repositories to make digitization easier in the first place. If validated across multi-center data, the approach could turn the world’s most routine blood test into a low-cost early warning system, catching anemia, infections and more serious hematological conditions before expensive, time-consuming confirmatory tests become necessary.

Subject of Research: AI-based multi-disease prediction from digitized complete blood count reports

Article Title: AI-enabled multi-disease diagnosis using low-cost CBC reports

Article References: Bibi, M., Saeed, W. K., Jabeen, S., Shah, S., Ali, G., ELAffendi, M., & Khan, M. (2026). AI-enabled multi-disease diagnosis using low-cost CBC reports. Frontiers in Imaging, 5, Article 1754419. https://doi.org/10.3389/fimag.2026.1754419

Image Credits: AI Generated

DOI: 10.3389/fimag.2026.1754419

Keywords: complete blood count, artificial intelligence, machine learning, OCR, PaddleOCR, multi-label classification, disease diagnosis, medical imaging, healthcare, anemia, random forest, gradient boosting

The research team, led by Shuang-shuang Zhang and colleagues spanning institutions in China and Italy, assembled an unusually large dataset for this kind of problem. Between January 2020 and December 2022, 12,105 infants underwent DDH screening at Shenzhen Children’s Hospital, each receiving a Graf-method hip ultrasound, the current gold standard for early detection, alongside a standardized physical examination performed by an orthopedic surgeon with more than twenty years of experience. Of those infants, 207, roughly 1.7 percent, were diagnosed with DDH, a prevalence consistent with the wide geographic and ethnic variation reported in the literature, which ranges from about 0.1 percent in African populations to 7.6 percent among Native Americans.

Raw data of this kind presents a classic statistical trap. Because DDH prevalence falls sharply as infants age, a naive model trained on the full cohort would learn to associate older age with lower risk, essentially memorizing the age distribution rather than the biology. The researchers countered this with age-stratified undersampling, drawing 11,898 DDH-negative infants down to match the age profile of the 207 positive cases. They then generated two balanced training cohorts, one with equal numbers of positive and negative cases and another with twice as many negatives, and split each into training and held-out test sets at an 8:2 ratio with fixed random seeds for reproducibility. Hyperparameter tuning and ten-fold cross-validation happened exclusively on the training data, keeping the test set blind until final evaluation.

Before any machine learning began, the team ran a clinical prescreening. Eighteen candidate variables went through univariate chi-square and t-tests, followed by multivariate logistic regression, a deliberate constraint to keep model complexity low given the modest number of positive cases and to anchor the final model to predictors clinicians already recognize. Six variables fell away, including birth weight, winter birth, cesarean delivery, and multiple births, leaving twelve: infant age, female sex, first-degree family history of DDH, premature birth, firstborn status, breech presentation, oligohydramnios, swaddling, frog-leg positioning, skin-fold asymmetry, restricted hip abduction, and leg length discrepancy. The logistic regression itself was revealing. A first-degree family history carried an odds ratio of nearly 42, by far the largest effect, while restricted hip abduction and leg length discrepancy each raised odds roughly tenfold or elevenfold. Frog-leg positioning or vertical holding, conversely, was protective, halving the odds, a finding that aligns with the mechanical theory that hips held in extension are deprived of the positioning that drives socket development.

Six algorithms competed for the top spot: random forest, support vector classification, logistic regression, k-nearest neighbors, LightGBM, and XGBoost, a gradient-boosted decision tree method known for both accuracy and speed. On the 1:1 balanced data, XGBoost won decisively, posting an area under the receiver operating characteristic curve of 0.883, accuracy of 0.807, precision of 0.744, an F1 score of 0.800, sensitivity of 0.865, and specificity of 0.761. The choice of the 1:1 cohort over the 1:2 cohort was not arbitrary. Models trained on the 1:2 data achieved higher specificity, peaking at 0.910, but their sensitivity topped out at 0.750, because the extra negative cases pushed the algorithms toward predicting negative outcomes. Given that a missed DDH diagnosis carries serious consequences for hip development, the researchers deliberately favored sensitivity, accepting more false positives in exchange for fewer missed cases.

The most striking result came from external validation. The team tested the XGBoost model on 80 infants screened at Jinhua Municipal Central Hospital in 2024, a completely separate institution, with 40 positive and 40 negative cases. Far from degrading, performance improved: the AUC rose to 0.909, accuracy to 0.838, precision to 0.829, and the F1 score to 0.840, with sensitivity of 0.850 and specificity of 0.825. This kind of cross-hospital transferability is the acid test for clinical prediction models, many of which quietly overfit to the quirks of a single dataset. The authors are careful to frame this as preliminary evidence, noting that the validation sample was small and drawn from a tertiary hospital, but the direction of the result is encouraging.

What elevates this work above a black-box classifier is its interpretability layer. The team applied SHAP, Shapley additive explanations, a technique borrowed from cooperative game theory that quantifies exactly how much each variable pushes an individual prediction up or down. Across the model, five variables dominated: asymmetry of thigh and groin skin folds, restriction of hip abduction, swaddling, breech presentation, and female sex. Two of the top five are pure physical examination findings, which the authors read as a directive: hands-on clinical assessment during infant checkups deserves at least as much attention as the demographic checklist. The SHAP summary also confirmed that swaddling, which holds the hips in prolonged extension, actively increases predicted risk, reinforcing a growing body of evidence, including a randomized controlled trial of traditional Mongolian swaddling, that holding practices shape hip development.

The model’s positioning within the broader artificial intelligence landscape is worth noting. Most AI research on DDH targets the diagnostic stage itself, using deep learning to read ultrasound images or pelvic radiographs automatically. Those systems require imaging equipment and trained operators. This model operates upstream, before any imaging, integrating demographic risk factors with bedside findings to produce a continuous, individualized risk estimate. Compared with earlier risk tools, such as a logistic regression calculator that reached an AUC of 0.81 without external validation, or a risk-factor counting rule that refers infants for ultrasound when two or more factors are present, the XGBoost approach offers graded prioritization rather than a binary threshold, potentially valuable where ultrasound capacity is limited.

The authors are candid about limitations. The cohort consisted of infants whose parents actively sought screening, introducing potential selection bias, and the retrospective design precludes causal inference. The 207 positive cases constrained how complex the models could be, and infants with syndromic, pathological, paralytic, or teratologic dislocations were excluded, so the model cannot be generalized to those subpopulations. A subtle methodological caveat involves the variable prescreening, which was performed on the full cohort before train-test splitting, a practice that can leak information; nested cross-validation with fold-specific feature selection would have been more rigorous but was impractical given the sample size. Perhaps most importantly, because both the internal test sets and the external validation set were artificially balanced, the predicted probabilities do not reflect the true 1.7 percent population prevalence. The authors explicitly recommend treating the model as a relative-risk ranking tool for prioritizing ultrasound referrals, not as a source of calibrated absolute risk, and they call for prospective cohorts with natural prevalence to evaluate calibration and decision-curve performance.

Even with those caveats, the study sketches a plausible future for infant hip screening. A clinician armed with twelve routinely collected data points, from family history to the symmetry of skin folds, could generate an individualized risk score in seconds and reserve scarce ultrasound slots for the infants most likely to need them. If prospective, multicenter validation confirms what this preliminary evidence suggests, the humble physical examination, augmented by gradient-boosted trees and game-theoretic explanations, may become the first line of defense against a condition that, untreated, follows children into adulthood.

Subject of Research: Machine learning prediction of developmental dysplasia of the hip in infants using clinical risk factors

Article Title: Machine learning–based personalized prediction of developmental dysplasia of the hip using risk factors and clinical examination

Article References: Zhang, S.-S., Caggiari, G., Chen, X.-Y., Huang, B.-X., Deng, H.-S., Xu, N., & Shi, W. (2026). Machine learning–based personalized prediction of developmental dysplasia of the hip using risk factors and clinical examination. Frontiers in Digital Health, 8, Article 1912106. https://doi.org/10.3389/fdgth.2026.1912106

Image Credits: AI Generated

DOI: 10.3389/fdgth.2026.1912106

Keywords: developmental dysplasia of the hip, machine learning, XGBoost, SHAP, infant screening, hip ultrasound, pediatric orthopedics, risk prediction, clinical examination, Graf method, gradient boosting, digital health

A new study published in Frontiers in Systems Biology tackles that bottleneck head-on with a biophysics-informed computational framework that predicts ASO knockdown efficacy from sequence data alone, even when training data are scarce, messy, and scattered across decades of heterogeneous literature. The work, led by Risha Guha with Shawn Kim, Anwita Bandaru, and Nithika Vivek, curates efficacy measurements from peer-reviewed publications and patent disclosures targeting ALS-relevant genes including SOD1, ATXN2, C9ORF72, FUS, AR, and STMN2, then demonstrates that a carefully engineered model can transfer its knowledge to entirely new target genes with as little as one experimental measurement per family.

The core insight of the framework is that sequence alone does not determine whether an ASO will work. A chemically perfect 20-mer is useless if its target site on the RNA is buried inside a stable secondary structure. The researchers therefore augmented conventional sequence features, such as GC content, CpG density, homopolymer runs, and regional composition of the 5-10-5 gapmer architecture, with physics-based descriptors computed using the ViennaRNA suite. Nearest-neighbor thermodynamic calculations via RNAduplex estimate the free energy of RNA-DNA hybrid duplex formation, while RNAplfold computes the energetic cost of opening the target site, a quantity the authors combine into an effective binding proxy reflecting both duplex stability and structural accessibility.

This mechanistic feature augmentation paid off decisively. When models trained on literature-derived data were tested against eight independently generated, quality-controlled experimental ASOs targeting the stress-granule-associated genes DAZAP1 and FAM69C, sequence-only baselines showed substantial external prediction error. Adding RNA-DNA hybridization thermodynamics produced the largest single improvement in accuracy on the external benchmark, and incorporating accessibility information improved generalization further. The result suggests that the first-order determinants of RNase H-mediated knockdown are thermodynamic in nature, and that explicitly encoding them allows a small-data model to generalize in ways purely statistical approaches cannot.

But a stubborn problem remained. Even with mechanistic features, uncalibrated predictions exhibited systematic family-specific scale bias. Predictions for DAZAP1 and FAM69C clustered within a narrow band even though observed knockdown spanned a much wider range, meaning the model preserved relative ordering within a family while missing the absolute level. The authors diagnose this not as a failure of learning but as missing information about family-specific baselines driven by differences in transcript structure, assay context, and cellular environment that no sequence-level model could capture.

Their solution is elegantly minimal: one-shot family calibration. For each new target gene, a single experimentally measured reference ASO provides an additive offset, computed as the difference between observed and predicted knockdown, which is then applied uniformly to all predictions within that family. No retraining, no additional complexity, no change to within-family ranking. On the experimental benchmark, this single adjustment dramatically reduced mean absolute error while leaving Spearman rank correlation stable, confirming that the model had learned transferable sequence-biophysical relationships and simply lacked a family-specific intercept.

To validate the calibration strategy beyond a handful of wet-lab points, the team ran a leave-one-family-out simulation on literature-derived ASO families, repeating the one-shot procedure one thousand times per family with randomly selected calibrators. The results were nuanced and honest: some families showed consistent error reduction, while others with high internal heterogeneity saw mixed or even negative effects. The authors emphasize this conservative behavior as a feature, not a flaw, because it means calibration does not artificially inflate accuracy when its assumptions fail, and supports selective rather than indiscriminate application.

The experimental side of the study is itself notable. Thirty candidate ASOs were transfected into SH-SY5Y cells using lipofection, RNA was extracted after five days, and knockdown was quantified by RT-qPCR with careful primer efficiency curves, outlier exclusion, and no-template controls. After quality control, six DAZAP1 and two FAM69C ASOs were retained, with the DAZAP1 analysis applying the Pfaffl method to correct for suboptimal primer amplification efficiency. The best DAZAP1 ASOs achieved roughly 22 percent transcript knockdown, while FAM69C ASOs reached 40 to 50 percent, providing genuine out-of-distribution test data that had never touched the model during training.

Interpretability analyses add a biological coda. Permutation-based feature importance revealed that global sequence architecture, summarized by principal components of 5-mer and 6-mer composition, dominates the predictive signal, with thermodynamic and accessibility features contributing correlated but secondary information. Partial dependence plots showed smooth, biologically plausible relationships, including threshold-like behavior for accessibility features consistent with the requirement that target sites exceed a minimum openness for productive RNase H binding. The pattern suggests that much of the relevant biophysics is already embedded in holistic sequence composition once correlated descriptors are compressed into orthogonal axes.

The limitations are acknowledged candidly. The dataset remains small, qualitative literature labels had to be mapped coarsely onto a numeric scale, knockdown was measured at the transcript level rather than the protein level, and chemical modifications and delivery efficiency were deliberately excluded to reduce confounding. Performance estimates for FAM69C, with only two experimental points, carried wide uncertainty intervals. Yet the framework offers something rare in computational drug discovery: a realistic workflow in which literature-trained models can be anchored to a new gene with a single wet-lab measurement, bridging the gap between globally trained predictors and the experimental reality of understudied therapeutic targets. For a field where every avoided screening round saves months and substantial cost, that is a meaningful step toward faster, smarter ASO design for ALS and beyond.

Subject of Research: Machine learning prediction of antisense oligonucleotide knockdown efficacy for ALS gene targets under small-data constraints

Article Title: Biophysics-informed computational framework for predicting antisense oligonucleotide knockdown efficacy under small-data constraints

Article References: Guha, R., Kim, S., Bandaru, A., & Vivek, N. (2026). Biophysics-informed computational framework for predicting antisense oligonucleotide knockdown efficacy under small-data constraints. Frontiers in Systems Biology, 6, Article 1921636. https://doi.org/10.3389/fsysb.2026.1921636

Image Credits: AI Generated

DOI: 10.3389/fsysb.2026.1921636

Keywords: antisense oligonucleotides, ALS, RNase H, machine learning, knockdown prediction, biophysics, RNA accessibility, one-shot calibration, gradient boosting, RT-qPCR, neurodegenerative disease, drug discovery

The study enrolled 101 patients newly diagnosed with metastatic colorectal cancer at a single academic center, all of whom were beginning first-line chemotherapy. The cohort had a median age of 62 years, and 69 percent of the patients were male. For each patient, the researchers compiled a rich set of clinical variables, ranging from demographic characteristics such as age at diagnosis to disease-specific features including the maximum dimension of metastatic lesions and whether metastases were present at the time of the initial diagnosis. The goal was deceptively simple: using only information available at or near the start of treatment, could an algorithm predict the best response a patient would achieve during first-line chemotherapy? This question matters enormously in clinical practice, because patients who are unlikely to benefit from a toxic and burdensome regimen might be spared unnecessary side effects, while those likely to respond could be prioritized for aggressive treatment.

To answer it, the team cast a wide net across the machine learning landscape. They evaluated fifteen distinct models, spanning the full spectrum of algorithmic families used in biomedical prediction research. Among them were classical statistical classifiers such as logistic regression, linear discriminant analysis and quadratic discriminant analysis, alongside naive Bayes methods that rely on probabilistic assumptions of feature independence. They also tested instance-based learning with K-nearest neighbors, margin-based classification with support vector machines, and a suite of ensemble approaches that combine many weaker learners into a stronger one: decision trees, random forests, extra trees, AdaBoost and gradient boosting classifiers. Each algorithm learns patterns from the training data in a fundamentally different way, and the researchers wanted to know which family of methods best captured the complex, nonlinear relationships between patient characteristics and treatment outcome.

The standout performer was LightGBM, the Light Gradient Boosting Machine, a relatively recent algorithm developed by Microsoft researchers that has become a favorite in tabular data competitions and biomedical applications alike. In its initial configuration, LightGBM achieved an accuracy of 0.71 in predicting chemotherapy response, outperforming the fourteen competing models. After the researchers tuned its hyperparameters, the model’s accuracy climbed to 0.79, with an area under the receiver operating characteristic curve, or AUC, of 0.82. These numbers are worth unpacking. Accuracy of 0.79 means the correctly classified roughly four out of five patients in the training data, while an AUC of 0.82 indicates that the model had a strong ability to discriminate between responders and non-responders across all possible classification thresholds. In oncology, where predictions guide decisions with life-and-death consequences, an AUC above 0.80 is generally considered a meaningful level of discriminative power.

LightGBM’s advantage stems from two clever engineering innovations that make gradient boosting both faster and less memory-hungry. Gradient boosting works by building an ensemble of decision trees sequentially, with each new tree trained to correct the errors of its predecessors. Traditional implementations evaluate every data point at every possible split point in every tree, which becomes computationally expensive as datasets grow. LightGBM introduces Gradient-based One-Side Sampling, or GOSS, which keeps data points with large gradients, meaning those the current model handles poorly, while randomly sampling from points with small gradients, since those are already well fit. It also employs Exclusive Feature Bundling, or EFB, which merges mutually exclusive sparse features into single features, dramatically reducing the number of dimensions the algorithm must consider. Together, these techniques allow LightGBM to handle clinical datasets efficiently while capturing subtle interactions between variables that simpler models would miss.

Perhaps the most clinically intriguing finding concerns which variables the model deemed most important. Feature importance analysis revealed that age at diagnosis, the maximum metastatic dimension, and metastatic status at diagnosis were the dominant predictors of chemotherapy response. In other words, the model leaned heavily on basic clinical and radiological characteristics rather than molecular data. Notably, genetic variables did not significantly impact the response predictions. This is a striking result given the central place that molecular biomarkers, such as RAS and BRAF mutation status, occupy in contemporary colorectal cancer decision-making. The finding may reflect the limited sample size, the specific set of genetic variables available in this cohort, or the possibility that, within this dataset, gross clinical features carried more predictive signal than the genetic information captured. Whatever the explanation, it suggests that inexpensive, routinely collected clinical data may hold more predictive power than clinicians might assume.

But the story takes a cautionary turn when the researchers evaluated their tuned model on a held-out test dataset, data the model had never seen during training. Performance dropped sharply: accuracy fell from 0.79 to 0.60. The authors interpret this decline as evidence of potential overfitting, the phenomenon in which a machine learning model memorizes the idiosyncrasies of its training data rather than learning generalizable patterns. Overfitting is an ever-present danger in medical machine learning, and it is particularly acute when the number of patients is small relative to the number of variables, as is the case here with 101 patients. A model with hundreds of tunable parameters can effectively memorize the outcomes of individual patients in the training set, producing impressive internal metrics that evaporate when confronted with new cases. The honest reporting of this performance drop is itself noteworthy, as the field of clinical machine learning has been criticized for publishing inflated results based solely on training or cross-validation performance.

The implications for clinical practice are therefore measured. The authors conclude that their LightGBM-based model demonstrated moderate performance in estimating response to first-line chemotherapy in metastatic colorectal cancer patients, but they are explicit that its clinical applicability remains limited and requires validation in larger, independent cohorts. This is the crucial next step for any predictive model aspiring to enter the clinic: external validation on data from different institutions, different populations and different time periods. A model that performs well only at the center where it was trained is of little use to the broader community of oncologists and patients. Larger multi-institutional datasets would also allow the models to better learn the true relationships between predictors and outcomes, potentially reducing the overfitting that undermined this study’s test performance and enabling more robust incorporation of genetic variables.

Despite these limitations, the study adds to a growing body of evidence that machine learning can extract meaningful predictive signals from routine clinical data in oncology. The identification of age at diagnosis, maximum metastatic dimension and metastatic status at diagnosis as key features aligns with clinical intuition and provides a starting point for future model refinement. The finding that fifteen different algorithms could be systematically compared on the same clinical question also illustrates how the methodological toolkit of data science is maturing within cancer research. As electronic health records accumulate richer data and computational methods grow more sophisticated, the vision of personalized treatment selection, in which an algorithm helps an oncologist choose the chemotherapy regimen most likely to work for a specific patient, moves closer to reality. This study, with its transparent reporting of both promise and failure, represents an honest step along that path rather than a premature declaration of victory.

For patients with metastatic colorectal cancer, the immediate takeaway is not that an algorithm will soon choose their treatment, but that the data needed to build such tools already exist in ordinary clinical records. The challenge ahead is assembling cohorts large enough, and diverse enough, to train models that generalize beyond a single hospital’s walls. If future studies can replicate and improve upon the discriminative power achieved here while eliminating the overfitting, gradient boosting machines and their kin could become standard instruments in the oncologist’s arsenal, helping to spare patients futile toxicity and direct them toward therapies with the greatest chance of benefit. The Turkish team’s work demonstrates both that this goal is within reach and that the road to trustworthy clinical artificial intelligence demands rigor, humility and, above all, validation on patients the model has never met.

Subject of Research: Machine learning prediction of first-line chemotherapy response in metastatic colorectal cancer

Article Title: A gradient boosting machine-based approach to predicting chemotherapy response in patients with metastatic colorectal cancer

Article References: Yıldız, O., Gürbüz, A. F., Eryılmaz, M. K., Araz, M., Yıldırım, M. S., Bozcuk, H. Ş., & Artaç, M. (2026). A gradient boosting machine-based approach to predicting chemotherapy response in patients with metastatic colorectal cancer. BMC Cancer. https://doi.org/10.1186/s12885-026-16347-x

Image Credits: AI Generated

DOI: 10.1186/s12885-026-16347-x

Keywords: metastatic colorectal cancer, machine learning, LightGBM, gradient boosting, chemotherapy response, predictive modeling, oncology, overfitting, clinical prediction, feature importance, BMC Cancer, precision medicine

The research, conducted by Sheetal Thomas of Parul University, Manjeevan Seera of Monash University Malaysia, and Mridula Goel of BITS Pilani, K. K. Birla Goa Campus, drew on data from more than three hundred participants recruited through universities, factories, government organisations and corporate offices across four districts of Gujarat, India. Each participant completed a sixty-item personality questionnaire covering the five major domains of the Big Five framework — Neuroticism, Extraversion, Openness to Experience, Agreeableness and Conscientiousness — and copied a standard passage known as the London Letter, a text deliberately chosen because it contains a broad range of letters, words and punctuation that makes visual characteristics comparable across writers.

The handwriting specimens were then translated into numbers. A coder with roughly twelve years of specialist handwriting-analysis training, working under the Hilliger Assessment Guidelines supplied by the British Institute of Graphologists, recorded between sixteen and twenty-six binary features for each personality domain. Slant, size, spacing, pressure, margins, baseline direction, letter forms and connecting strokes were each scored as present or absent. The personality outcomes, meanwhile, were computed as continuous scores: the mean of twelve five-point Likert items per domain, yielding values on a scale from one to five rather than the binary personality labels that have complicated much of the earlier literature.

What makes this study unusual is not the data itself but the rigor of its evaluation. The researchers first audited the source workbooks, excluding sixty-two pilot-phase records and screening out incomplete, invalid or invariant questionnaires within each trait separately. That produced analytic samples ranging from 320 to 330 participants per domain. They then compared two quite different machine-learning models — elastic-net regression, a regularised linear method that shrinks weak coefficients, and histogram-based gradient boosting, a non-linear ensemble of constrained regression trees — against a deliberately humbling benchmark: a mean-prediction baseline that ignores the handwriting entirely. Every model was evaluated with five-fold cross-validation repeated five times, generating twenty-five out-of-sample test folds for each model and domain combination.

The results were unambiguous. Neither model consistently improved on the mean baseline in mean absolute error, the study’s primary metric. For Neuroticism, the baseline error was 0.3102 score units while elastic net managed 0.3104 — a difference in the fourth decimal place. Extraversion and Openness told the same story. Agreeableness was the single domain where elastic net edged out the baseline, but the advantage amounted to roughly 0.0011 score points, about one-thousandth of a single point on the five-point scale. Gradient boosting actually performed worse than the baseline in every domain. Most tellingly, mean cross-validated R² was negative for every model-domain combination, meaning that in held-out data the models predicted worse than simply predicting the training-set average.

Permutation tests reinforced the conclusion. When the researchers randomly shuffled the outcome values fifty times and re-evaluated each model, none of the observed model performances stood out from what chance would produce. The smallest permutation p-value was 0.078, for elastic net on Agreeableness, and the rest ranged far higher. The authors are careful to note that these exploratory tests used a limited number of shuffles, but taken together with the negative R² values and the negligible error reductions, the pattern points firmly in one direction: the handwriting-feature matrices contained no reliable information about the questionnaire-derived scores in this sample.

Perhaps the most interesting twist lies in what did work. Two independent, practising graphologists, blinded to all questionnaire data and to each other’s ratings, separately coded twenty-six common handwriting samples. They agreed with one another to a striking degree: mean feature-level percentage agreement ranged from 94.1 percent to 97.8 percent across domains, and mean Cohen’s kappa — a statistic that corrects for chance agreement — ranged from 0.803 for Neuroticism to 0.929 for Conscientiousness. In other words, the features could be identified reproducibly. Yet this reliability did nothing to establish validity. As the authors emphasise, inter-rater reliability and criterion validity answer different questions: one asks whether coders see the same thing, the other whether that thing predicts an independent measure of the construct. A feature can be recognised with near-perfect consistency and still carry no psychological meaning.

The study also probes the quality of the reference measures themselves. Internal consistency of the study-specific questionnaire varied substantially, with Cronbach’s alpha of 0.466 for Neuroticism and 0.675 for Openness — below the commonly cited 0.70 threshold — against 0.805 for Conscientiousness. The authors acknowledge that measurement error of this kind can attenuate any true association and cap the achievable predictive performance, so the low reliabilities limit how strongly the negative result can be generalised. But they are equally clear that weak questionnaire reliability does not rescue handwriting-based claims; it simply constrains conclusions in both directions. The study establishes that these specific coded features did not predict these specific scores, not that no conceivable handwriting representation could ever relate to a better-validated personality instrument.

The context of earlier work matters here too. Parts of the same Gujarat dataset had been used in four previous publications, covering image-based detection of Extraversion- and Neuroticism-associated features, sentiment and risk profiling from handwritten text, and questionnaire-based links between personality and mutual-fund investment preferences. The authors stress that this new paper is not a replication or repudiation of those studies but a different analytical task with a stricter estimand: predicting an independently constructed continuous personality score from the complete feature matrix, out of sample. Earlier feature-detection results, they argue, should not be read as demonstrations that handwriting can recover continuous personality scores in unseen individuals without further evidence.

The broader implications reach well beyond graphology. Personality has been linked in the behavioural-finance literature to risk tolerance, investment preferences and planning behaviour, and that association creates commercial temptation to score personality from cheap, easily collected behavioural traces — including handwriting — in finance, employment, insurance and other high-stakes settings. This study demonstrates why such shortcuts are dangerous. A technically reproducible coding pipeline feeding a properly implemented machine-learning model can still fail completely at the criterion it claims to measure, and algorithmic output lends an unwarranted air of objectivity to constructs that lack validated measurement. The authors recommend independent, preregistered, externally validated evidence, subgroup fairness analyses and clear incremental predictive value before any such tool touches consequential decisions. In publishing a transparent negative result with complete data and code — the coded features are archived on Harvard Dataverse and the full Python workflow is released — the team offers a template for how behavioural AI claims should be tested: clear outcome construction, a simple baseline, repeated out-of-sample evaluation and the honesty to report when the signal simply is not there.

Subject of Research: Predictive validity of manually coded handwriting features for questionnaire-derived Big Five personality scores

Article Title: Predictive validity of manually coded handwriting features for questionnaire derived personality scores

Article References: Thomas, S., Seera, M., & Goel, M. (2026). Predictive validity of manually coded handwriting features for questionnaire derived personality scores. Discover Artificial Intelligence, 6(1), Article 1489. https://doi.org/10.1007/s44163-026-02132-z

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02132-z

Keywords: graphology, handwriting analysis, Big Five personality, predictive validity, machine learning, elastic net, gradient boosting, cross-validation, inter-rater reliability, negative result, psychometrics, behavioural AI

The most common response to incomplete records is deceptively simple: throw them away. Analysts who delete every row containing a missing entry, an approach known as complete-case analysis, end up with a dataset that is smaller, potentially biased and statistically less powerful. If the people who failed to answer a questionnaire differ systematically from those who did, the surviving sample no longer represents the population it was drawn from. Even when the missingness is benign, discarding data inflates uncertainty and weakens statistical estimates, a cost that grows rapidly as the proportion of missing entries climbs.

Imputation offers an alternative. Instead of deleting incomplete cases, imputation methods replace each missing value with a plausible estimate inferred from the rest of the data. Done well, this preserves sample size, retains information that deletion would destroy and yields more reliable downstream estimates. The catch is that imputation is itself a prediction problem: to fill in a missing number, an algorithm must model how that number relates to everything else in the record. The quality of the fill-in therefore depends entirely on the quality of the predictive model doing the filling, which is precisely where the new study concentrates its effort.

Asif and Samart’s proposal, Layered Stacking, is built on the idea of stacking, an ensemble strategy in which multiple different models are trained to solve the same problem and their predictions are then combined by a higher-level model that learns when to trust each one. Rather than relying on a single imputation algorithm, Layered Stacking organizes several base learners into layers, letting each layer refine the estimates produced by the one beneath it. The architecture is designed specifically for continuous missing values, the numeric fields where small imputation errors can compound into distorted means, variances and regression coefficients.

To find out whether the added complexity pays off, the researchers ran comprehensive simulation studies built on four distinct data-generating processes. Each simulated dataset was generated under the missing-at-random mechanism, the widely assumed scenario in which the probability that an entry is missing depends on other observed variables but not on the missing value itself. The experiments spanned sample sizes of 500, 1,000, 5,000 and 10,000 observations, and missingness rates of 10, 20, 30 and 40 percent, a grid that captures everything from small studies with modest gaps to large datasets with substantial holes.

The benchmark competition was demanding. Layered Stacking was pitted against five imputation techniques that are widely used in practice: k-nearest neighbors, which averages the values of the most similar complete records; support vector regression, which fits flexible boundaries in a transformed feature space; decision trees, which partition the data through a cascade of yes-or-no questions; random forests, which average many decorrelated trees; and extreme gradient boosting, the sequential ensemble method that has dominated tabular machine-learning competitions for years. Each method was asked to impute the same missing entries under the same conditions, allowing a direct, like-for-like comparison.

Performance was scored with two complementary metrics. The mean squared imputation error, or MSIE, penalizes large mistakes heavily and therefore rewards methods that rarely produce wildly wrong estimates. The mean absolute percentage imputation error, or MAPIE, expresses the average error as a percentage of the true value, making results comparable across variables measured on different scales. The authors also reported the standard deviation of the errors across replications, along with 95 percent confidence intervals, so that differences between methods could be judged against the noise inherent in repeated simulation runs rather than a single lucky or unlucky draw.

The headline result is that Layered Stacking generally achieved lower MSIE than all five benchmark models, with its advantage becoming particularly pronounced on larger datasets. That pattern makes intuitive sense: ensemble methods that learn how to combine base learners need enough data to estimate the combination reliably, so the payoff of stacking grows as sample size increases. For analysts working with the large, messy datasets that characterize contemporary research, the finding suggests that the extra computational cost of a layered ensemble may be repaid in measurably more accurate imputations, which in turn feed into more trustworthy statistical models built on top of the completed data.

Simulations alone can flatter a method in ways real data never do, so the authors also tested Layered Stacking on two real-world datasets. These datasets bring irregularities that no data-generating process fully captures: skewed distributions, awkward correlations between variables and missingness patterns shaped by human behavior rather than mathematical rules. The real-data evaluation served as a sanity check on the simulation findings, and the study reports that the method’s performance carried over, reinforcing the case that the gains are not an artifact of carefully controlled synthetic conditions.

The broader significance of the work lies in what it says about the state of imputation practice. Methods like k-nearest neighbors and random forest imputation have become default choices in statistical software, and extreme gradient boosting has been increasingly adopted for its raw predictive power. Showing that a layered stacking ensemble can beat these established baselines, especially at scale, nudges the field toward treating imputation not as a preprocessing afterthought but as a modeling problem worthy of the same ensemble sophistication applied to prediction tasks. For anyone who has watched a regression analysis wobble because a fifth of the data contained holes, the message is appealing: the blanks in a dataset need not be a permanent tax on statistical accuracy, and the tools for filling them are getting better.

Subject of Research: Machine learning methods for imputing missing continuous data

Article Title: A layered stacking ensemble approach for missing data imputation

Article References: Asif, M., & Samart, K. (2026). A layered stacking ensemble approach for missing data imputation. PLOS One, 21(10), e0359843. https://doi.org/10.1371/journal.pone.0359843

Image Credits: AI Generated

DOI: 10.1371/journal.pone.0359843

Keywords: missing data, imputation, stacking ensemble, machine learning, PLOS One, simulation study, missing at random, random forest, gradient boosting, statistics, data science, mean squared error