VitaAI
Role: Team Lead : ML prediction pipeline, MERN backend architecture, module interconnection, end-to-end system design
Overview
Doctors make life-or-death decisions under time pressure with incomplete information. VitaAI is an AI-assisted medical diagnosis platform built for doctors, not patients. Input a patient's symptoms, and the system predicts the top 10 most likely diseases ranked by severity, generates SHAP-based clinical reasoning for each prediction, and translates that reasoning into a plain-language medical report via IBM WatsonX. Doctors can generate prescriptions directly. Patients can schedule appointments and access their reports. Validated and tested end-to-end by 16 real-world doctors. Won the Deloitte Capstone Project Award.
The Problem
A junior doctor spends years learning to do one thing well: look at a set of symptoms, rank the possible diagnoses by likelihood and severity, and explain the clinical reasoning behind each. Most hospitals don't have enough junior doctors. Most clinics can't afford them at all. VitaAI was built to fill that gap, not to replace doctors, but to give every doctor a tireless, explainable AI assistant that does the differential diagnosis groundwork instantly, so the doctor can focus on the decision rather than the derivation.
How It Was Built
Differential diagnosis — ranking possible diseases against a patient's symptoms — is one of the most cognitively demanding tasks in medicine. A doctor must simultaneously consider dozens of conditions, weigh symptom combinations, account for severity, and articulate reasoning clearly enough to document and defend. Junior doctors spend years developing this skill. VitaAI compresses that process into seconds, with full explainability.
Crucially, VitaAI was not built for patients. It was built for doctors. A system that hands a patient a list of ten possible diseases creates anxiety and misuse. A system that hands a doctor a ranked, reasoned differential with clinical language they can act on creates value.
The Three-Stage Pipeline
A doctor enters a patient's symptoms through an interactive chatbot interface. The system runs three stages in sequence.
| Stage | Name | What Happens |
|---|---|---|
| 1 | Disease Prediction | XGBoost classifier predicts top 10 most likely diseases, ranked by severity-weighted probability — not raw probability alone |
| 2 | SHAP Clinical Reasoning | SHAP explainer generates feature attribution for each predicted disease — which symptoms contributed positively, which worked against it, and by how much |
| 3 | WatsonX Narrative Report | SHAP values passed to IBM WatsonX, which converts quantitative reasoning into a clinically worded, documentable narrative for each prediction |
Stage 1 — Severity-Weighted Ranking
The ranking is not purely by probability. A rare but life-threatening condition with moderate symptom overlap will rank higher than a common but benign one. This is the correct clinical prioritisation: a doctor needs to rule out the worst thing first.
Stage 2 — SHAP as Clinical Reasoning
SHAP attribution is the quantitative equivalent of a junior doctor saying *"I'm thinking sepsis because of the fever and elevated heart rate, but the absence of localised pain makes it less certain."* The same logic — made consistent, reproducible, and auditable.
Stage 3 — Narrative Output
The doctor receives not just a ranked list but a readable, documentable explanation for every prediction — the kind of clinical summary they would write themselves, generated in seconds. From here they can generate a prescription directly on the platform and share reports through the patient-facing portal.
The Patient Interface
The patient-facing interface is intentionally minimal. Patients can schedule appointments, view doctor-generated reports, and access prescriptions. They cannot see raw diagnostic predictions — that information belongs in a clinical context, not a consumer one.
Clinical Validation — 16 Real Doctors
This is the detail that separates VitaAI from most medical AI projects built in academic settings.
Before launch, prediction outputs were shared with 16 practising doctors. They reviewed top-10 predictions across a range of symptom combinations, provided feedback on ranking accuracy and clinical plausibility, and their input was used to align the model to real-world clinical judgment. Those same 16 doctors then tested the full system end-to-end.
| Validation Step | Detail |
|---|---|
| Model review | 16 doctors reviewed top-10 predictions across multiple symptom combinations |
| Feedback loop | Ranking corrections and clinical plausibility feedback incorporated into model |
| End-to-end testing | All 16 doctors tested the full flow: symptom entry → SHAP report → prescription |
| Outcome | Final system reflects clinical feedback at every layer |
This is not a model trained on a dataset and shipped. It is a model trained, validated, corrected, and endorsed by people who actually practice medicine.
Tech Stack
| Layer | Technology |
|---|---|
| Frontend | React — separate doctor and patient interfaces |
| API Gateway | Node.js + Express.js — orchestrates frontend, Python ML backend, and WatsonX API |
| Database | MongoDB — user data, appointments, prescriptions, report storage |
| Auth | JWT-based — separates doctor and patient access scopes |
| ML Backend | Python — XGBoost inference, SHAP computation, WatsonX report generation (modular service) |
| LLM | IBM WatsonX |
| Explainability | SHAP |
The ML pipeline is fully decoupled from the API and frontend — model updates require no changes to either layer.
Recognition
VitaAI won the Deloitte Capstone Project Award, selected from competing teams across the institution. Judges specifically cited the clinical validation approach and the explainability architecture as the distinguishing factors.
Results & Impact
Won the Deloitte Capstone Project Award. Prediction model validated, boosted, and alignment-tested by 16 real-world doctors across end-to-end testing.