A multimodal risk stratification model using deep learning and tabular data

Cardiovascular disease and stroke kill millions every year, yet medicine still waits for the first symptom before acting. A team of computer scientists from Ain Shams University in Cairo has now built a deep learning model that flips that logic on its head, reading an ordinary table of clinical measurements from a single doctor’s visit and estimating a patient’s risk of both conditions at once. The work, published in Discover Artificial Intelligence, reports an accuracy of 88.83 percent and an area under the receiver operating characteristic curve (AUROC) of 0.9124 on its main test set, and it held up surprisingly well when confronted with completely independent data from the UK Biobank. The researchers say the goal is not to replace clinicians but to hand them a stratification tool that works from a cross-sectional snapshot, without longitudinal records, imaging, or pre-computed risk scores.

The scale of the problem the model targets is hard to overstate. According to the World Health Organization, cardiovascular diseases were responsible for an estimated 19.8 million deaths in 2022, roughly 32 percent of all global mortality, and cerebrovascular disease claimed another 6.6 million lives in the same year. Eighty-five percent of these deaths occur in low- and middle-income countries, where diagnostic infrastructure is often limited and reactive practice is the norm. Many patients are diagnosed only after a myocardial infarction or an ischemic stroke, stages at which treatment outcomes are markedly less favorable. The Egyptian team argues that this reactive paradigm is precisely what a well-designed predictive model can disrupt, identifying high-risk individuals before an acute event forces the issue.

Technically, the framework rests on a merged dataset of 71,000 samples and 16 harmonized features. The researchers combined the Kaggle Heart Disease Dataset, drawn from the UCI Machine Learning Repository and containing 1,000 real cardiac patient records, with a Stroke Risk Prediction Dataset of approximately 5,110 real records. Because deep learning is data-hungry and the stroke data was scarce, they expanded the stroke cohort to 70,000 samples using a Conditional Tabular GAN, or CTGAN, a generative adversarial network trained exclusively on the 5,110 real samples to preserve statistical distributions and inter-feature correlations. Before merging, the team harmonized the two feature spaces: the stroke dataset’s binary high blood pressure flag was mapped to the heart dataset’s continuous resting blood pressure using the clinical threshold of 140 mmHg or greater, chest pain was encoded into four ordinal severity categories, and smoking status and sex were standardized to uniform binary encodings. A pre-computed stroke risk percentage was strictly excluded to prevent target leakage.

Two architectures were evaluated. The baseline was a deliberately lightweight Multi-Layer Perceptron built in PyTorch, comprising exactly three parametric layers: an input projection of dimension 16, a single hidden layer of 64 neurons with ReLU activation and Dropout regularization at a rate of 0.2, and a dual-output layer with Sigmoid activation producing independent probability estimates for cardiovascular disease and stroke. The proposed model was an adapted TabNet, an attention-based architecture optimized for tabular data, reconfigured for multi-task learning. TabNet’s shared encoder runs through five sequential decision steps, each computing an attention mask over all 16 input features with sparse regularization, and the learned masks are accumulated across steps to yield global feature importance scores. Two independent linear output heads, one for cardiovascular disease and one for stroke, attach at the final decision step, so task-specific differentiation emerges from the shared encoder’s attention allocation rather than from task-specific hidden layers.

Class imbalance, a persistent headache in medical datasets, was attacked from two directions. Upstream, the team applied SMOTE-NC oversampling, a variant of the Synthetic Minority Over-sampling Technique designed for mixed numerical and categorical features, but crucially only to the post-split training partition, ensuring no synthetic minority samples contaminated the validation or test sets. Downstream, they used Focal Loss with alpha set to 0.25 and gamma set to 2.0, computed independently per output head and summed with equal weighting. Focal Loss down-weights easy majority-class examples, forcing the network to concentrate gradient updates on hard-to-classify minority instances. The data pipeline itself was engineered for integrity: CTGAN synthesis used no test-set records, min-max normalization parameters were fitted solely on training data, and hyperparameter tuning occurred on an inner validation set of 5,680 samples before a single unbiased evaluation on the held-out 14,200-sample test set.

The results tell a clear story. The MLP baseline achieved a macro-averaged accuracy of 79.66 percent with an AUROC of 0.8213, precision of 0.7803, recall of 0.7500, and an F1-score of 0.7600. The multi-task TabNet reached 88.83 percent accuracy, an AUROC of 0.9124, precision of 0.8736, recall of 0.8541, and an F1-score of 0.8673, a 9.17 percent absolute improvement in accuracy with non-overlapping confidence intervals confirming statistical significance. An ablation study isolated the value of multi-task learning itself: single-task TabNet variants achieved CVD and stroke AUROCs of 0.897 and 0.886 respectively, while the shared encoder lifted these to 0.921 and 0.904. The improvement was most pronounced for cardiovascular disease, the minority task, consistent with established findings that lower-resource tasks benefit most from shared representations.

Interpretability, often the Achilles heel of deep learning in medicine, was addressed through a dual-layer approach. TabNet’s built-in attention mechanism revealed that the cardiovascular head prioritized resting blood pressure, contributing 25 percent of the attention-weighted decision process, and chest pain type at 20 percent, while the stroke head prioritized the hypertension flag at 30 percent and age at 22 percent. Crucially, these task-specific profiles diverge despite sharing the same encoder, confirming that the attention mechanism allocates different feature subsets per disease rather than collapsing to a single weighting. SHAP value decomposition corroborated the rankings at the local level, and in a representative high-risk cardiovascular case, elevated blood pressure, asymptomatic chest pain, and advanced age emerged as the dominant positive contributors, a pattern fully consistent with ACC/AHA cardiovascular risk criteria. The team also assessed calibration using Brier scores, with TabNet scoring 0.094 for cardiovascular disease and 0.087 for stroke, and Platt scaling reduced maximum bin-level miscalibration to 0.02 for both tasks.

The most consequential test was external validation on the UK Biobank, a prospective database of approximately 500,000 participants. The researchers selected a subset of 10,000 participants aged 40 to 75 with complete data for all 16 features and independently recorded ICD-10 diagnoses, comprising 5,823 males and 4,177 females with a mean age of 56.4 years, cardiovascular disease prevalence of 18.3 percent, and stroke prevalence of 9.7 percent. On this entirely real, unseen cohort, TabNet achieved an AUROC of 0.85, with 0.83 for cardiovascular disease and 0.87 for stroke, and a macro-averaged accuracy of 83.1 percent. This figure aligns closely with an internal real-data-only evaluation, which yielded an AUROC of 0.847 on the combined 6,110 authentic records, and the agreement between the two provides the study’s strongest evidence that the model captures genuine clinical signal rather than artifacts of synthetic data.

That synthetic dependence is the study’s most candid limitation, and the authors confront it directly. Because 98.6 percent of the training samples were CTGAN-generated, the team ran a sensitivity analysis across augmentation levels, observing monotonic AUROC improvement from 0.847 on real data only to 0.891 and then 0.912 as synthetic volume increased, suggesting the generator provides useful distributional coverage rather than noise. They also acknowledge an asymmetry in label provenance: the stroke label is directly observed, while the cardiovascular label for stroke-origin samples is a clinically grounded proxy derived from established risk thresholds. The UK Biobank validation, where cardiovascular outcomes rest on ICD-10 diagnoses entirely independent of the labeling rule, bounds the impact of this design choice. Preliminary experiments traced a near-perfect stroke AUROC of 0.99991 to target leakage from a derived feature, and its exclusion reduced the score to a realistic 0.9043, a transparency the researchers present as a cautionary tale for the field.

The implications reach well beyond the benchmark numbers. A model that stratifies risk from a single clinical visit snapshot, running on an ordinary CPU with 16 gigabytes of RAM, could be deployed in primary-care settings where imaging modalities and longitudinal electronic health records are unavailable, and where 85 percent of the global cardiovascular and stroke burden actually falls. The dual-task design mirrors clinical reality, since hypertension, obesity, and age drive both conditions simultaneously, and it delivers not just a risk score but a task-specific explanation of which factors drive each disease independently. The authors propose extending the framework to additional chronic conditions such as diabetes, integrating multimodal inputs like ECG signals or wearable sensor data, and prioritizing cohorts where both outcomes are independently recorded for the same patients. They also stress that ethical considerations, particularly the mitigation of demographic and socioeconomic biases, must remain central as such models move toward clinical decision support. If the approach generalizes as the UK Biobank results suggest, the reactive era of cardiovascular medicine may finally be giving way to something predictive, auditable, and within reach of the clinics that need it most.

Subject of Research: A multimodal risk stratification model using deep learning and tabular data

Article Title: A multimodal risk stratification model using deep learning and tabular data

Article References: Sadik, A. A., Morsey, M. M., Nazmy, T., & El-Horbaty, E.-S. M. (2026). A multimodal risk stratification model using deep learning and tabular data. Discover Artificial Intelligence, 6(1), Article 1218. https://doi.org/10.1007/s44163-026-02225-9

Image Credits: AI Generated

DOI: 10.1007/s44163-026-02225-9

Keywords: multimodal, risk, stratification, model, deep, learning, tabular, data, scientific research