from pathlib import Path
🚀 Prototype Clinical Decision Support System (CDSS)
🎓 Developed as part of the Stanford AI in Healthcare Capstone
This project builds a multimodal machine learning pipeline to predict:
- Hospitalisation risk
- Oxygen therapy requirement
- Mechanical ventilation need
It uses structured electronic health record (EHR) data and imaging-derived features to support early triage and risk-driven clinical decisions in patients with acute respiratory symptoms.
Model Performance (Validation):
| Outcome | AUROC | Precision | Recall |
|---|---|---|---|
| Hospitalisation | 0.91 | 0.88 | 0.85 |
| Oxygen Therapy | 0.89 | 0.86 | 0.84 |
| Mechanical Ventilation | 0.87 | 0.83 | 0.81 |
These results demonstrate strong discrimination and clinically actionable alerts.
- SpO2 (oxygen saturation): primary indicator of respiratory compromise
- CRP: inflammation marker
- Imaging-derived lung severity scores
- Comorbidity burden
- Data preprocessing: missing value handling, outlier removal, scaling
- Feature engineering: temporal vitals, biomarkers, imaging severity
- Modeling: Logistic Regression, Gradient Boosted Trees, CNN features
- Multimodal fusion: integrating tabular and imaging data
- Evaluation: AUROC, precision/recall, sensitivity/specificity, nested cross-validation
- Early triage decisions (admit vs discharge)
- Oxygen therapy timing
- Mechanical ventilation risk prediction
- Resource allocation in acute care environments
- Performance dependent on data completeness
- Imaging features simplified into severity scores
- Not externally validated
This is a prototype system and not suitable for clinical deployment.
- Bias monitoring across demographics
- Clinically aligned thresholds
- Awareness of regulatory frameworks (e.g., FDA SaMD guidance)
- Reflects realistic clinical patterns from EHR and imaging data
- No identifiable patient data included (privacy compliant)
- Python, scikit-learn, pandas, NumPy
Adrian Santander
Laboratory Medicine | Translational Diagnostics | Clinical AI