Development of an AI system for disease risk prediction
Standard screenings miss up to 40% of cases in early stages — a patient with three risk factors receives intervention when the disease has already developed.
We build AI models that analyze medical history, genetics, and lifestyle: the system indicates risk 2–5 years before clinical manifestations.
For cardiovascular risk, the model achieves AUC 0.86 — 15% higher than classic SCORE2, and due to calibration, the Brier Score decreases by 20%.
Preventive medicine stops being an abstraction — for 100k patients, the system can identify up to 2,000 high-risk individuals and prevent hundreds of hospitalizations, saving an annual amount calculated after audit.
Disease risk prediction tasks
Population screening — identifying high-risk individuals across the entire assigned population. Use cases: type 2 diabetes, CVD, oncology, chronic kidney disease. Individual prediction — 10-year risk of a cardiovascular event.
ML models outperform classic risk scores (Framingham, SCORE2) by 15–20% in AUC (recent meta-analysis) due to nonlinear interactions and a larger number of predictors. Disease progression — a patient with an early stage: when will it progress to severe? We use survival models (Cox PH, Random Survival Forest, DeepHit) with time-to-event endpoints.
How does an AI model account for genetic data?
Genomic data — SNPs for polygenic risk scores. Polygenic risk score (PRS) is a weighted sum of thousands of SNPs. The ML task: optimal weighting for a specific population. In our projects, we combine PRS with EHR data: a multimodal model yields an AUC improvement of 5–8%.
Why is calibration more important than accuracy?
A risk score of "68%" should mean exactly a 68% probability. After training, we apply Platt scaling or isotonic regression. Calibration plot (reliability diagram) is a mandatory metric. Without calibration, a model can be confidently wrong, which is clinically dangerous.
What data sources are used?
- Structured EHR data: diagnoses (ICD-10), laboratory tests (glucose, HbA1c, lipids), medication prescriptions, vital signs, demographics.
- Genomic data: SNPs for polygenic scores. BRCA1/2, ApoE4, PCSK9.
- Lifestyle and social factors: smoking, BMI, diet, stress, education level. From EMR, questionnaires, wearable devices.
| Data type | Examples | Role in model |
|---|---|---|
| EHR | Diagnoses, laboratory tests, prescriptions | Main predictors, temporal dynamics |
| Genomic | SNPs, PRS | Additional signal for hereditary risks |
| Lifestyle | Smoking, BMI, activity | Modifiable factors, improve calibration |
How we build models: stack and approach
- Preprocessing: missing value imputation, diagnosis encoding, time series aggregation.
- Feature engineering: creating combined features (e.g., Charlson comorbidity index, time windows of laboratory values).
- Training: XGBoost or LightGBM — for tabular data with SHAP interpretation. For time series — Transformer-based models (BEHRT, Med-BERT) with pretraining on EMR.
- Calibration: Platt scaling or isotonic regression.
- Validation: temporal split and external validation on data from another clinic.
Comparison: XGBoost + SHAP gives a 1.5x AUC improvement compared to logistic regression on the same data, and due to calibration, the Brier Score decreases by 20%.
Models and validation
Technical validation details
For tabular EHR data, we use XGBoost and LightGBM — they dominate on real-world medical data. Advantages: handling missing values, interpretability via SHAP, efficiency on small samples. For time series — Transformer-based models (BEHRT, Med-BERT). Pretraining on EMR databases → fine-tuning for specific risk tasks.Evaluation metrics
| Metric | Clinical meaning |
|---|---|
| AUC-ROC | Discrimination: separates sick from healthy |
| AUC-PR | Under strong class imbalance (rare events) |
| Brier Score | Overall accuracy of probabilistic predictions |
| Net Benefit / Decision Curve | Clinical utility at threshold decisions |
| NRI, IDI | Improvement vs. existing risk score |
External validation on data from another clinic is mandatory before implementation. Implementation in population health: stratification by risk score — high risk → active outreach (call, screening invitation). In the EMR, the score is displayed with a SHAP explanation: "CVD risk: 23% (high).
Factors: hypertension, dyslipidemia, smoking." Economic effect: in a population of 100k people, we identify 1,500–2,000 high-risk individuals → intervention prevents 200–400 hospitalizations, yielding net annual savings calculated after audit.
What is included in the work
- Analysis of the clinic's data and EMR structure.
- Development of a predictive model (XGBoost/Transformer) for a specific outcome.
- Calibration and external validation.
- Integration of the risk score into the EMR (REST API + SHAP widget).
- Documentation, physician training, 6-month post-release support.
Our team's experience — 5+ years in medical AI, 30+ projects in prediction and EHR processing. We guarantee methodological rigor: all models undergo external validation and calibration.
Get a consultation on implementation: we will send examples of implemented risk score systems under NDA. Contact us to evaluate your project.
