💗 VitalPulse AI — Cardiovascular Risk Intelligence

Your health, decoded. Your risk, explained. Your care, connected.

Elevator Pitch: AI-powered cardiovascular risk intelligence platform that predicts, explains, and acts — with 15 health modules, specialist booking, live ECG monitoring, and one-click clinical documents.

💡 Inspiration

Every year, 17.9 million people die from cardiovascular disease — not because the science to catch it early doesn't exist, and not because the warning signs are absent. But because the systems meant to help people act on those warnings are deeply, dangerously broken.

What hit us hardest was not the number itself — it was the pattern behind it:

🧑‍⚕️ A 52-year-old man leaves a clinic with a "moderate risk" label — no explanation of what's driving it, no roadmap to change it, no way to measure if anything he does helps. He walks out with anxiety but without direction.
👩‍⚕️ A woman managing diabetes needs a retinal screening, a cardiology consult, and a specialist referral — each on a separate platform, a separate appointment system, a separate stack of forms. Every handoff loses critical information. Days stretch into weeks. The window for early intervention narrows.
🩺 A primary care physician seeing 30 patients a day spends 20 minutes manually writing each handoff brief — documentation time that should have been spent in the room with the next patient.

None of these are medical failures. They are infrastructure failures. And infrastructure failures are exactly what software is built to solve.

We built VitalPulse AI because we believed one well-designed, integrated intelligence platform could close that gap — not partially. Entirely. A platform that doesn't stop at prediction, but explains, forecasts, acts on, and monitors that prediction over time.

🚀 What It Does

VitalPulse AI is an end-to-end cardiovascular intelligence and comprehensive health platform. At its core, a deep learning model trained on 56,000+ real biomedical records analyzes patient vitals and returns a cardiovascular risk probability classified as Low, Moderate, or High.

But the score is only the beginning. The platform delivers a complete intelligence loop across six seamlessly connected stages:

🔍 Predict → 💡 Explain → 📈 Forecast → ⚡ Act → 📊 Monitor → 🛡️ Audit

🩺 Real Stories, Real Solutions

Arjun, 48 — Catching the Silent Risk Before It Becomes a Crisis

Enters his vitals → receives 23.1% cardiovascular risk probability (Low)
SHAP-powered chart flags blood pressure and physical inactivity as top drivers
Cardio Intelligence Brief generates instantly: reduce sodium, add 20 min walking daily, schedule a lipid panel
Downloads patient report, shares with GP → first time he has direction, not just a number

Priya — Seamless Specialist Access When It Matters Most

Navigates to Find Your Specialist → filters by Cardiology → finds Dr. Nabila Rahman (4.9★, $210)
Reviews Monday/Wednesday slots → books Wednesday 10:00 in a single click
Instant confirmation on screen — no phone calls, no waiting rooms, no third-party redirects

Marcus — Retinal Screening from Home

Opens Diabetic Retinopathy module → uploads retinal scan image
Receives severity classification in seconds: Mild DR, 78% confidence
Downloads result, books ophthalmology consult — all within the same session

Dr. Chowdhury — VitalPulse as a Clinician's Handoff Tool

System flags 74.5% risk, identifies systolic BP and glucose as primary drivers
Generates a phased intervention roadmap
One click → Provider Handoff Brief exported in under 60 seconds (vs. 20 minutes manual)

Sasha — Monitoring 30 Days of Real Progress

Baseline risk: 61% → Day 15 current status: 54% → Target: 45%
Adherence sitting at 82% — she can watch the changes working in real time

🎛️ The Integrated Intelligence Dashboard

The command center of the entire platform — no tab-switching, no separate logins, no lost context. Platform-wide trust anchors displayed at a glance: 🔐 5L Encrypted · 24/7 Monitoring · 99% Accuracy · AI-Powered.

📡 Clinical Signal Command Center — live frequency data, risk distribution signals, triage alerts, and forecast confidence updating in real time (synchronized badge indicator)
🔔 Smart Symptom Triage Engine — accepts symptoms + intensity levels → returns urgency classification: Stable / Attention / Urgent with recommended follow-up window and routing action (e.g. Chest Pain + High intensity → Stable → "Run Full Health Check" recommended)
💊 Daily Adherence Monitor — tracks MyFitnessPal goals, Medication Reminder, 10k Steps Daily, and Hydration Plans with visual completion checkboxes
🧘 AI Posture Coach — real-time posture feedback with instant visual correction guidance
🏋️ Smart Workout Coach — personalized session plans based on current health data, displayed as a widget alongside the posture coach
💓 Live ECG Monitor (IoT) — streams a continuous real-time waveform with automatic anomaly detection, live status badge, and AI Analysis Status: Normal Sinus Rhythm auto-detection label — live cardiac rhythm with no additional hardware required
⚡ Clinical Ops Shortcuts — one-click panel to instantly: Open Memo Studio · Launch Health Check · Book a Specialist · Explore Services — zero navigation friction from the dashboard

🧠 Cardiovascular Risk Analysis Engine

Inputs accepted: Age, gender, height, weight, systolic & diastolic blood pressure, cholesterol, glucose, smoking status, alcohol use, physical activity level.

Internally computes: BMI, pulse pressure, mean arterial pressure, age group classification, composite risk score.

Output delivered:

🎯 Probability score with Low (<40%) / Moderate (40–70%) / High (>70%) classification
📊 Visual risk gauge for immediate comprehension
📉 Key Risk Factors bar chart — top contributing variables + impact magnitude (SHAP-powered)
💬 Personalized AI recommendation tailored to the exact patient profile
📁 Timestamped result logged into rolling risk trajectory for longitudinal monitoring

Every prediction powered by SHAP-based explainability — reveals not just what the risk is, but precisely why, across every health marker simultaneously.

📋 Cardio Intelligence Brief

Converts each prediction into a structured, actionable clinical plan:

🔢 Confidence Score — model certainty on the prediction (e.g. 81% confidence on a 23.1% CVD probability result)
📉 Risk Reduction Estimate — projected % reduction achievable through intervention (e.g. Risk Score 5.63 → target reduction path)
🚦 Priority Mode Signal — Preventive Maintenance / Active Management / Urgent Review (e.g. "Preventive Maintenance Mode — Next Review Within 72h")
📏 Care Momentum Meter — adherence tracked visually
🗂️ Phased Action Roadmap — organized into Immediate Actions / This Week / This Month

🧪 7-Day Scenario Lab

Projects four distinct intervention paths simultaneously with real projected deltas:

Path	Projected Risk	Delta	Impact
📊 Baseline	23.1%	0.0%	Reference
💊 BP Optimization	~20.4%	-3.7%	Moderate
🏃 Lifestyle Upgrade	~23.1%	~0.0%	Low
🌟 Comprehensive Plan	~17.1%	-6.6%	Maximum ✅

Best-performing scenario highlighted automatically — Comprehensive Plan projects the maximum -6.6% delta from baseline
Each path shows projected risk value, delta from baseline, and impact classification with color coding

📅 30-Day Prevention Outcome Tracker

Makes progress of interventions measurable and visible day by day:

Baseline risk → Current status → Target risk → Adherence % — all in a single dashboard
Turns prevention from an abstract goal into a trackable, motivating reality

📄 One-Click Clinical Documents

With a single click from any prediction:

🧾 Patient Report — structured, timestamped, formatted for personal use
🏥 Provider Handoff Brief — fully structured clinical document ready for specialist referral
⏱️ What used to take 20 minutes of manual documentation takes under 60 seconds

🤖 AI Health Co-Pilot — 24/7 Intelligence at Every Page

A persistent intelligent assistant woven into every page — always context-aware:

💬 Conversational health guidance — cardiac reports, risk levels, medication interactions, general questions answered in natural language
🎤 Voice input via browser-native speech recognition — completely hands-free
📸 Screen Capture Analysis — capture current screen, ask the Co-Pilot to explain or act on what it sees
📎 PDF & Image Upload — medical reports, lab results, retinal scans dropped directly into chat
From within the Co-Pilot: book urgent specialist appointments, export handoff briefs, have responses read aloud for accessibility

🏥 15 Specialized Health Modules

Module	What It Does
🏃 Fitness Tracker	Tracks calories burned, optimizes workout efficiency
💧 Diabetes Check	Predicts diabetes risk from glucose, insulin, BMI with SHAP
❤️ Heart Health	Comprehensive cardiovascular screening using key vitals
🧠 Parkinson's Detection	Voice analysis markers + neurological feature extraction
🫁 Lung Health	Assesses lung cancer risk from lifestyle data + symptom input
🔬 Thyroid Function	Evaluates hypothyroid conditions from hormonal profile data
🏋️ Exercises	Scientifically backed, goal-based workout plans per health status
🥗 Diet Plans	Personalized nutrition strategies with macro analysis + dietary targets
📐 BMI Calculator	Body composition baseline with health risk context + clinical insights
🤖 AI Nutritionist	Conversational real-time meal + nutrition guidance
🩹 Injury Prevention	Recovery strategies + sport-specific prevention protocols
📹 Live Exercise	Real-time pose tracking with rep counting + automatic form correction
👨‍⚕️ Expert Consultation	End-to-end video consultation booking with certified specialists
🌸 Women's Health	Cycle tracking, fertility monitoring, and pregnancy care
👁️ Diabetic Retinopathy	AI retinal image screening — classifies No DR / Mild / Moderate / Severe with confidence scores

👨‍⚕️ Specialist Discovery & Appointment Booking

Healthcare Services hub stats: 24/7 AI triage support · 12+ specialist categories · 1-min average booking flow

12 verified specialists across 9 medical fields — all bookable in under 60 seconds:

🫀 Cardiology · 🧠 Neurology · 🩺 Dermatology · 🔪 Surgery · 🦴 Orthopedics · 👶 Pediatrics · 🎗️ Oncology · 🌸 Gynecology · 🔬 Endocrinology

Each specialist profile carries a verified star rating (⭐4.5 to ⭐5.0), consultation fees ranging from $120 to $240, designated clinic location, and multilingual consultation support — spanning VitalPulse Central, VitalPulse Heart Institute, VitalPulse Children's Clinic, VitalPulse Metabolic Center, Mount Adora Hospital Sylhet, and dedicated specialty centers for Oncology, Women's Health, and Endocrinology.

Real-time search bar filtering by name or specialization
Every doctor card: photo, specialty badge, verified rating, consultation fee, clinic location
Full Doctor Profile reveals: education history with institutions + dates, career experience, current position, available time slots by day, services offered, multilingual support (e.g. Dr. Sarah Smith: English, Bangla, Hindi)
Direct booking panel: date selection → time slot selection → instant confirmation · free reschedule policy
Virtual Treatment Flow: Book consultations → Receive guided plans online or in-person → Track progress and return for follow-up sessions — everything stays in one track

3 Focused Care Programs:

💪 Conditioning — targeted routines to improve baseline strength and stamina
🤸 Mobility — joint-safe movement plans with guided form support
🔄 Recovery — progressive follow-up care for long-term health outcomes

Medical Services covered on platform: Cancer Care · Labor & Delivery · Heart & Vascular · Mental Health · Neurology · Burn Treatment

Patient Testimonials: Real verified feedback from platform users — Dr. Zakir (Cardiology Patient, 5★), Elba Martin (Preventive Care, 5★), Marco Silva (Neurology Patient, 5★) — displayed on the services hub with star ratings and direct quotes on care quality and booking experience.

📂 History, Tracking & Lifecycle Management

📋 Medical History Tab — complete log of every risk assessment across all 15 modules with date, prediction type, risk level, probability score, individual and bulk-delete controls
🏃 Workout History Tab — timestamped interactive timeline of all exercise sessions and fitness tracker results
📅 Appointments Tab — upcoming and past consultations with doctor name, date, time, approval/payment status, cancellation controls, new booking directly from tab
💳 Integrated Payment Simulation — processes appointment fees within the platform, no external payment dependency
⭐ Doctor Reviews & Dynamic Ratings — every new review automatically recalculates the doctor's live star rating in real time

📈 Model Performance & Safety Telemetry

All 8 algorithms benchmarked side by side on 56,000+ real biomedical records across every clinically relevant metric:

Model	Accuracy	ROC-AUC	F1	MCC	Kappa	Sensitivity	Specificity
GBM ⭐	0.735	0.801	0.725	0.472	0.470	0.698	0.772
Neural Network	0.732	0.799	0.712	0.468	0.464	0.664	0.800
DNN	0.730	0.795	0.715	0.463	0.460	0.678	0.783
Random Forest	0.730	0.794	0.719	0.461	0.460	0.690	0.769
LSTM	0.727	0.797	0.705	0.459	0.454	0.654	0.800
SVM	0.726	0.785	0.724	0.451	0.451	0.719	0.733
Adaboost	0.725	0.792	0.702	0.454	0.449	0.651	0.799
Logistic Regression	0.711	0.777	0.700	0.424	0.423	0.674	0.748

🏆 Top Performers: GBM leads on Best Accuracy (0.735) and Best ROC-AUC (0.801). An Ensemble model combining multiple algorithms typically provides the most robust predictions for deployment.

Three critical audit layers — built in from day one:

📈 Calibration Curve — validates model confidence is well-calibrated across the full probability range (when model says 40%, it means 40% in real-world outcomes)
⚖️ Fairness Slice Comparison — compares observed vs. predicted positive rates across Male, Age <45, Age 45–59, Age 60+ subgroups — full demographic disparity transparency with side-by-side bar charts
🔍 Drift Summary — per-feature drift monitoring in production: Overall 0.1564 · Age 0.2552 · ap_hi (systolic) 0.0701 · ap_lo (diastolic) 0.1472 · Cholesterol 0.2708 — detects distribution shifts before they silently degrade performance

Live Safety Telemetry: ROC-AUC 0.799 · Brier Score 0.183 · Sample Size 56,000 · Drift Status MODERATE

🌍 Multi-Language Support & Clinical Safety

The entire platform — risk outputs, clinical documents, AI Co-Pilot, all 15 modules — available in multiple languages
A patient in rural Bihar and a clinician in Boston can both use the platform fully in their preferred language with no degradation in intelligence or clinical depth
⚠️ VitalPulse AI is a clinical decision-support platform — for risk awareness, pattern detection, and workflow acceleration. It is not a replacement for licensed medical diagnosis, emergency care, or physician judgment. All high-risk outputs should be evaluated by a qualified healthcare professional.

🔬 How We Built It

Machine Learning Backbone:

🧠 TensorFlow / Keras — primary deep learning cardiovascular model
🌲 XGBoost, LightGBM, CatBoost, Random Forest, SVM, LSTM, Adaboost — all rigorously benchmarked on accuracy, ROC-AUC, F1, MCC, Kappa, sensitivity, specificity
⚙️ Optuna — hyperparameter optimization across hundreds of trial runs
⚖️ imbalanced-learn — addressed class imbalance to ensure reliability across the full risk spectrum
💡 SHAP + LIME — feature-level explainability on every single prediction
👁️ Separate computer vision pipeline for Diabetic Retinopathy retinal image classification

Backend:

🐍 Flask + Flask-SQLAlchemy — API and data persistence
🔐 PyJWT — session token authentication
📄 ReportLab + FPDF2 — clinical-grade PDF generation
🚀 Waitress — production WSGI server
🔧 Pandas, NumPy, Joblib, Pydantic — data transformation, model persistence, payload validation

Frontend:

🎨 Jinja2 + Bootstrap 5 + Tailwind CSS — layout and styling
📊 Plotly — all interactive clinical visualizations
🤖 Puter.js — AI Co-Pilot chat integration
📸 html2canvas — screen capture functionality
📄 PDF.js — in-browser document rendering

Infrastructure: SQLite database with locally stored model artifacts — every feature built from scratch and integrated into a single product. Nothing assembled from disconnected external services.

🧩 Challenges We Ran Into

⚡ Challenge 1: Making 15 Features Feel Like One Platform

Early versions had the prediction engine, scenario lab, appointment system, history tracker, and document generator as essentially separate mini-applications
Completely rethought the data architecture and user flow multiple times — designed every stage around the outputs of the previous one
Built shared data models that persist patient context across every module → the result was a fundamental change in what the platform could do in a real clinical moment

⚡ Challenge 2: Running SHAP at Real-Time Inference Speed

SHAP KernelExplainer is computationally expensive — early testing added several seconds to every prediction, making real-time clinical use impractical
Replaced KernelExplainer with TreeExplainer for tree-based models + careful background sample sizing
Restructured inference pipeline to compute SHAP values in the same forward pass as the risk prediction — eliminating redundant model load operations
Result: sub-second explanation latency — every bar in the Key Risk Factors chart loads alongside the risk score, not after it

⚡ Challenge 3: Genuine Fairness Auditing, Not Decorative Fairness Auditing

First fairness slice comparison run revealed real disparities between observed and predicted positive rates across age cohorts — not explainable by training set noise
Returned to the training pipeline, reexamined data collection and preprocessing decisions, retrained with corrected class weighting and calibration strategies
Treated the fairness slice comparison as a regression test every new model version must pass before deployment
The current model still shows a moderate drift score — an honest acknowledgment that no model is perfect and that real-world production data always diverges from training data

🏆 Accomplishments That We're Proud Of

✅ Genuinely complete platform — not a prototype with hollow features. Every module works. Every prediction is explainable. Every booking is real. Every clinical document generates correctly. The ECG streams. The retinopathy classifier classifies. The scenario lab simulates. All 15 modules — end to end.
📊 Model performance that tells a clinical story — GBM at ROC-AUC 0.801, accuracy 0.735, Brier Score 0.183 on 56,000+ records. Not numbers optimized for leaderboard aesthetics — numbers that reflect real-world calibration.
🛡️ Safety telemetry by design — calibration curves, fairness slice comparisons, and drift monitoring built in from the start. Clinical AI without accountability infrastructure is not a product. It's a liability.
🔄 The loop is genuinely closed — a patient can go from entering vitals → downloading a clinical document → booking a specialist → tracking 30-day intervention progress entirely within one session, on one platform, with no fragmentation and no information lost between steps.

📚 What We Learned

🧠 The hardest part of clinical AI is not the model — it's the trust layer around it. Accuracy alone is not enough. If a clinician can't understand why the model flagged a patient as high risk, the output is clinically unusable regardless of the ROC-AUC.
⚖️ Explainability, fairness auditing, and drift monitoring are not optional features for a future version — they are the baseline for responsible deployment and must be first-class engineering requirements from day one.
🔗 The most dangerous failures in healthcare technology are not technical failures — they are coordination failures. Fragmented platforms, manual documentation, disconnected referral pathways: none of these are technically hard. They are coordination problems. And coordination problems are exactly what well-designed integrated software solves.
🏥 Treat every feature decision as a clinical decision first and a software decision second. The question was never just can we build this — it was always what does a real patient or clinician actually need this to do at the moment it matters most.

🔭 What's Next for VitalPulse AI

⌚ Real Wearable Integration — replace the simulated ECG stream with genuine sensor ingestion from consumer health devices (smartwatches, biosensor patches), making live monitoring truly continuous and patient-driven
🏥 Clinician-Facing Cohort Analytics Panel — lets physicians see population-level risk trends across their full patient panels, flag at-risk cohorts proactively, and track intervention effectiveness at scale → shifts VitalPulse AI from individual decision support to population health intelligence
🧠 Longitudinal Adherence Scoring — a persistent model layer that learns from each patient's medication logs, activity data, and risk trajectory over months, dynamically adjusting intervention recommendations rather than producing a single static roadmap that goes stale
🌍 Multilingual Clinical Document Templates + Regional Care Pathway Routing — the same intelligence, the same accountability, the same continuity of care, in any language, in any region, at any point in the care journey
📡 A patient in rural Bihar and a clinician in Boston — the same platform, the same depth, the same outcomes

The infrastructure is built. The loop is closed. Now it scales.

VitalPulse AI is a clinical decision-support platform. It is not a replacement for licensed medical diagnosis, emergency care, or physician judgment. All high-risk outputs and urgent triage signals should be evaluated by a qualified healthcare professional.