MediSync

AI Powered Intelligent Clinical Triage and Decision Support Platform

MediSync is a production grade clinical decision support system (CDSS) built for emergency departments, urgent care centers, and outpatient clinics. It combines real time large language model reasoning with deterministic clinical safety rules to help physicians make faster, more accurate diagnostic and triage decisions.

The platform is designed to feel like something between OpenEvidence, Glass Health, and an AI powered emergency department triage tool. It does not replace doctors. Every prediction includes confidence scores, supporting evidence, and references to current clinical guidelines.

What MediSync Does

Instead of asking "what disease do I have," MediSync answers the questions that actually matter in clinical practice:

How urgently does this patient need treatment
What are the probable conditions and what evidence supports them
What tests should be ordered next
What interventions are appropriate
How confident is the AI in its assessment

The platform performs a full clinical analysis pipeline on every patient:

Patient understanding (demographics, symptoms, history, medications, allergies, vitals, labs, lifestyle)
Medical reasoning (LLM generates differential diagnoses with evidence)
Risk scoring (validated clinical decision rules)
Triage assignment (ESI v4 + MEWS inspired scoring)
Test recommendations (evidence based workup)
Treatment suggestions (guideline referenced)
SOAP note generation (LLM written clinical documentation)
ICD 10 coding (automatic)
CPT billing code suggestions
FHIR R4 interoperability export

Key Features

Real LLM Clinical Reasoning

MediSync uses the GLM 4 large language model via a standard OpenAI-compatible API to generate:

4 to 5 ranked differential diagnoses with ICD 10 codes, probability scores, and confidence levels
Per diagnosis clinical reasoning explaining what evidence supports and argues against each condition
A full 2 to 3 paragraph clinical reasoning narrative referencing specific guidelines
Real time chat responses with patient specific data and guideline citations
Complete SOAP notes (Subjective, Objective, Assessment, Plan) with patient specific values

The LLM calls are streamed using Server Sent Events with heartbeat bytes to prevent gateway timeouts. A robust JSON repair function handles common LLM output errors including missing colons, missing commas, and markdown code block wrappers.

Deterministic Clinical Safety Rules

The LLM does not handle clinical safety. These rules run instantly and deterministically:

Vital sign analysis against standard reference ranges (heart rate, blood pressure, SpO2, temperature, respiratory rate, blood glucose, BMI)
Laboratory value analysis with abnormality flagging (hemoglobin, WBC, platelets, sodium, potassium, creatinine, BUN, glucose, ALT, AST, troponin, CRP, ESR, INR, HbA1c)
Drug drug interaction checking against a curated database of 6 medications with mechanism, clinical effect, and management
Allergy cross reactivity checking (penicillin, sulfa, aspirin, peanut, latex)
Duplicate therapy detection
Triage scoring using ESI v4 principles and Modified Early Warning Score (MEWS)
Risk scores inspired by CURB 65, CHA2DS2 VASc, qSOFA, Wells criteria, and NIHSS
Hospital admission prediction (discharge, observation, admission, ICU)
30 day readmission risk estimation
30 day mortality risk estimation

Multi Patient Management

Add unlimited patients via the onboarding flow or the sidebar
Each patient has independent clinical data and analysis
Patients persist in localStorage across page reloads
Switch between patients instantly by tapping their initials block in the sidebar
Delete patients with a two step confirmation

Vitals Radar Visualization

A hexagonal SVG radar chart on the Overview screen that visualizes all 6 vital signs as a single shape:

Normal vitals produce a balanced hexagon
Abnormal vitals push the vertex inward (orange)
Critical vitals create a deep notch (red with pulsing animation)
A faint green dashed hexagon shows the normal reference shape
You can see at a glance whether a patient is sick based on the shape distortion alone

AI Chat Assistant

A conversational interface where physicians can ask:

What supports the leading diagnosis
Why was a specific test recommended
What are the red flags for this patient
Could this be a different condition
What does the latest guideline say

The chat references the patient actual clinical data and cites specific guidelines (ACC/AHA, SSC, AHA/ASA, ATS/IDSA, ADA, ESC, KDIGO) with evidence level and publication year.

SOAP Note Generation

One tap generates a complete clinical note:

Subjective: chief complaint, HPI, PMH, medications, allergies, social history
Objective: vital signs with specific values, lab abnormalities, exam findings
Assessment: diagnostic impression with leading differential and ICD 10 code
Plan: 5 to 7 numbered items covering workup, therapeutics, monitoring, disposition, follow up, patient education

Includes automatic ICD 10 coding and CPT billing code suggestions with RVU values.

What If Simulator

Adjust vital signs with sliders and see how the AI risk score responds in real time. The simulator uses clinical heuristics (tachycardia, hypoxemia, hypotension) to compute a simplified risk delta.

Knowledge Graph

An interactive SVG graph of 48 clinical entities (diseases, symptoms, drugs, tests, body systems, complications, treatments) connected by 60+ relationships. Click any node to explore its connections.

Technology Stack

Frontend

Next.js 16 with App Router and Turbopack
TypeScript 5 with strict typing throughout
Tailwind CSS 4 for styling
shadcn/ui component library (New York style) with Radix UI primitives
Framer Motion for animations and transitions
Recharts for data visualization
Zustand for client state management
Bangers display font for headlines
JetBrains Mono for body text and data labels

Backend

Next.js API Routes (App Router server components)
Server Sent Events for streaming LLM responses with heartbeats
Standard Fetch API for GLM 4 large language model access

Data

localStorage for patient persistence (no external database required)
In memory Zustand store for reactive state management
Curated medical knowledge base in TypeScript (ICD 10 codes, drug interactions, clinical guidelines)

Design

TypeUI Tetris design system with periwinkle surface, navy primary, purple secondary
Glassmorphism with backdrop blur on cards, sidebar, header, and dock
Floating macOS style dock for zone navigation (no navbar)
Hard offset shadows (3px 3px 0) for the Tetris block aesthetic
Light mode only (no dark mode)

Architecture Overview

Patient Data Entered
       |
       v
engine.ts (instant, deterministic)
  -> Vital sign analysis
  -> Drug interaction checking
  -> Allergy cross reactivity
  -> Triage scoring (ESI + MEWS)
  -> Risk scores (CURB 65, CHA2DS2 VASc, qSOFA)
  -> Admission prediction
       |
       v
/api/clinical-reasoning (streaming, 5 to 30 seconds)
  -> GLM 4 generates differentials + narrative
  -> JSON repair handles LLM output errors
  -> Merged with deterministic analysis
       |
       v
UI displays combined result
  -> "REAL LLM" badge if GLM 4 succeeded
  -> "RULE BASED" fallback if GLM 4 failed

The architecture is intentionally hybrid. Clinical safety rules (vital signs, drug interactions, triage) are deterministic because you cannot rely on an LLM to consistently catch a drug interaction or compute a triage score. The LLM handles the reasoning layer (differentials, narrative, chat, documentation) because that requires medical knowledge and natural language generation.

Real AI Integration

MediSync uses a direct API client which calls the GLM 4 large language model. This is a real hosted LLM API, not a local model, not a mock, not a simulation.

Three API Endpoints

POST /api/clinical-reasoning

Accepts patient data and returns a complete clinical analysis. The response is streamed using Server Sent Events with heartbeat bytes every 3 seconds to prevent gateway timeouts. The LLM generates 4 to 5 differential diagnoses with ICD 10 codes, probabilities, confidence, reasoning, evidence factors, red flags, and recommended tests. It also generates a 2 to 3 paragraph narrative. The deterministic engine runs first (instant) and the LLM output is merged on top.

POST /api/chat

Accepts a question, patient data, current analysis, and conversation history. Returns a real LLM response with reasoning trace and evidence citations from the guideline corpus. The system prompt includes the full patient context and 8 guideline documents for RAG.

POST /api/soap

Accepts patient data and analysis. Returns a complete SOAP note as JSON (subjective, objective, assessment, plan) generated by the LLM with patient specific values.

JSON Repair

LLMs occasionally produce malformed JSON. The repairLLMJSON() function handles:

Missing colons before arrays: "redFlags[" becomes "redFlags":[
Missing colons before objects: "keyName{" becomes "keyName":{
Missing commas between objects: }{ becomes },{
Missing commas between arrays: ][ becomes ],[
Single quotes converted to double quotes
Trailing commas removed
Markdown code block wrappers (triple backtick json) stripped
Unclosed code blocks handled

If parsing fails after repair, the system retries the LLM call up to 2 times before falling back to deterministic analysis.

Probability Normalization

LLMs sometimes return 0 to 1 fractions instead of 0 to 100 percentages. The normalizePercent() function detects values of 1 or less and multiplies by 100, then clamps to the 5 to 98 range.

Clinical Safety Rules

The deterministic engine (src/lib/medical/engine.ts) implements the following clinical rules:

Vital Sign Analysis

Each vital sign is checked against standard adult reference ranges with normal, low, high, critical low, and critical high classifications:

Vital Sign	Normal Range	Critical Low	Critical High
Heart Rate	60 to 100 bpm	40 bpm	130 bpm
Systolic BP	90 to 120 mmHg	80 mmHg	180 mmHg
Diastolic BP	60 to 80 mmHg	50 mmHg	110 mmHg
Temperature	36.1 to 37.2 C	35.0 C	39.5 C
Respiratory Rate	12 to 20 per min	8 per min	30 per min
SpO2	95 to 100 percent	88 percent	100 percent
Blood Glucose	70 to 140 mg/dL	50 mg/dL	400 mg/dL

BMI is calculated from weight and height and classified as underweight, normal, overweight, or obese.

Drug Interaction Checking

The system checks the patient medication list against a curated database of 6 drugs (Warfarin, Metformin, Lisinopril, Atorvastatin, Amoxicillin, Furosemide) with full interaction data:

Mechanism of interaction
Clinical effect
Management recommendation
Severity (contraindicated, major, moderate, minor)
Evidence level (A, B, C)

Duplicate therapy detection flags when two drugs from the same class are prescribed.

Allergy Cross Reactivity

Checks patient allergies against 5 allergen classes (Penicillin, Sulfa, Aspirin, Peanut, Latex) with cross reactive drug lists and alternative medications.

Triage Scoring

Uses a hybrid of ESI v4 principles and Modified Early Warning Score (MEWS):

SpO2 below 88 percent triggers red triage immediately
Systolic BP below 80 triggers red triage
Heart rate above 130 or below 40 adds significant points
Temperature above 40 C adds points
Pain score 9 or 10 adds points
Top differential probability above 50 percent in cardiovascular, neurological, or multi system adds points
Critical lab values (troponin, potassium) add points
Symptom severity modifies the score

Triage levels: Red (resuscitation, immediate), Orange (emergent, 1 hour), Yellow (urgent, 24 hours), Green (less urgent, 3 days), Blue (non urgent, 1 week).

Risk Scores

Six risk scores are computed:

Cardiovascular Disease (10 year ASCVD) using Pooled Cohort Equations inspired scoring
Ischemic Stroke using CHA2DS2 VASc inspired criteria
Sepsis (in hospital) using qSOFA inspired criteria
Type 2 Diabetes using ADA risk test and HbA1c stratification
Acute Kidney Injury using KDIGO creatinine criteria
30 Day Mortality using APACHE II inspired scoring

Each risk score includes contributing factors, risk level (low, moderate, high, very high), and a clinical recommendation.

Admission Prediction

Predicts disposition (Discharge, Observation, Admission, ICU) based on triage level, vital sign abnormalities, lab values, and critical findings. Also estimates length of stay and 30 day readmission probability.

Medical Knowledge Base

The knowledge base (src/lib/medical/knowledge.ts) contains:

Disease Database

10 ICD 10 coded diseases with:

Body system classification
Prevalence and mortality rates
Key symptoms
Red flags
Recommended tests
First line treatment
Guideline references

Diseases included: Acute Coronary Syndrome, Community Acquired Pneumonia, Sepsis, Ischemic Stroke, Type 2 Diabetes, Acute Pyelonephritis, Acute Gastroenteritis, Acute Pulmonary Embolism, Hypertensive Emergency, Chronic Kidney Disease.

Drug Database

6 drugs with full interaction matrices:

Warfarin (anticoagulant)
Metformin (biguanide)
Lisinopril (ACE inhibitor)
Atorvastatin (statin)
Amoxicillin (beta lactam antibiotic)
Furosemide (loop diuretic)

Each drug includes class, half life, indications, contraindications, interactions with mechanism and management, and pregnancy category.

Clinical Decision Rules

Compiled criteria for:

CURB 65 (pneumonia severity)
CHA2DS2 VASc (stroke risk in atrial fibrillation)
qSOFA (sepsis screening)
Wells Criteria (pulmonary embolism)
NIHSS (stroke scale)

RAG Knowledge Corpus

8 real clinical guideline excerpts used by the LLM chat for evidence based citations:

ACC/AHA 2022 Chest Pain Guidelines
Surviving Sepsis Campaign 2021
AHA/ASA 2019 Acute Ischemic Stroke Guidelines
ATS/IDSA 2019 Community Acquired Pneumonia Guidelines
ADA Standards of Care 2024
ESC 2019 Pulmonary Embolism Guidelines
KDIGO 2024 Chronic Kidney Disease Guidelines
AHA 2017 Hypertension Guidelines

Knowledge Graph

48 nodes (diseases, symptoms, drugs, tests, body systems, complications, treatments) connected by 60+ weighted edges representing clinical relationships (presents with, diagnosed by, treated by, complicates to, affects).

ICD 10 and CPT Codes

ICD 10 codes for 25+ common diagnoses with descriptions. CPT codes for 14 common procedures with RVU values for billing.

User Interface

Design System

MediSync uses the TypeUI Tetris design system:

Surface: Periwinkle (#DFE7FF) with a subtle gradient mesh background
Primary: Deep navy (#1C202B) for borders, headings, and structure
Secondary: Vivid purple (#7107E7) for accents, active states, and LLM badges
Success: Green (#16A34A) for normal vitals and low risk
Warning: Orange (#D97706) for abnormal vitals and moderate risk
Danger: Red (#DC2626) for critical alerts and high risk
Display font: Bangers (bold, game inspired)
Body font: JetBrains Mono (monospace, technical)
Cards: Frosted glass (rgba(255,255,255,0.75) with backdrop blur)
Shadows: Hard offset (4px 4px 0) for the Tetris block aesthetic plus soft drop shadows for depth
Corners: 4px radius (almost square, game like)

Navigation

MediSync has no navbar. Zone navigation is done through a floating macOS style dock at the bottom center of the screen:

4 zone icons (Overview, Patient, Reasoning, Tools) with labels
Hover effect: icons lift up 6px and scale 1.1x with spring animation
Active zone: highlighted purple with white dot indicator
Glassmorphic: dark glass with backdrop blur, purple border, rounded corners
On mobile: labels hide, dock shrinks to icons only

Layout

Desktop (768px and up):

Left sidebar (240px, fixed, dark glass) with patient directory
Main content area (margin left 240px) with glassmorphic header and content
Floating dock at bottom center

Mobile (below 768px):

No visible sidebar (hidden off screen)
Hamburger menu in header opens sidebar as slide in drawer with overlay
Full width content
Compact floating dock (icons only)

Onboarding Flow

When the app loads with no patients, a full screen intake form appears:

Bangers headline "WELCOME TO MEDISYNC"
Complete clinical intake: demographics, symptoms (with quick add chips), duration, severity, vitals (6 fields), key labs (6 fields), medications, allergies, medical history, smoking toggle
Submit button disabled until name and at least one symptom are entered
On submit: patient is added, app transitions to main interface, AI analysis auto runs

Zones

01 Overview

Hero patient card with triage verdict
Leading diagnosis with probability and confidence
Vitals radar (hexagonal SVG that distorts with abnormalities)
Quick action buttons (numbered 01 to 04)
Risk score summary (4 stat blocks)

02 Patient

Demographics form
Symptom entry with quick add chips
Vital signs editor (6 fields with pain score slider)
Key labs editor (6 fields with abnormality indicators)
Medications, allergies, and medical history managers
Lifestyle factors (smoking, alcohol, exercise, diet)

03 Reasoning

Triage banner with level and disposition
Ranked differentials list (click to select)
Selected diagnosis detail with probability, confidence, reasoning, evidence bars, and red flags
Full LLM generated narrative
Recommended workup grid (up to 9 tests with priority badges)

04 Tools

Tabbed interface with 4 tools:
- AI Assistant (chat with LLM, patient specific, guideline citations)
- SOAP Note (LLM generated, copy and download, ICD 10 and CPT codes)
- What If (vital sign sliders with real time risk delta)
- Knowledge Graph (interactive SVG, click to explore)

Patient Workflow

Onboarding: User enters patient data (name, age, symptoms, vitals, labs, medications, allergies, history)
Analysis: The deterministic engine runs instantly, then the LLM generates differentials and narrative (5 to 30 seconds)
Overview: User sees the clinical snapshot with triage verdict, leading diagnosis, vitals radar, and risk scores
Reasoning: User reviews the ranked differentials, evidence, red flags, and full narrative
Tools: User can chat with the AI about the case, generate a SOAP note, simulate what if scenarios, or explore the knowledge graph
Documentation: User generates a SOAP note with ICD 10 and CPT codes, copies or downloads it
Multi Patient: User can add more patients, switch between them via the sidebar, and each patient retains independent analysis

Getting Started

Prerequisites

Node.js 18 or higher
Bun (for package management and running)
An API key is required to power the clinical reasoning features.

API Key Configuration

Configure your environment variables in .env (or .env.local) in the root of the project to set your API key and base URL:

ZAI_API_KEY=YOUR_Z_AI_API_KEY
ZAI_BASE_URL=https://api.z.ai/api/paas/v4

Installation

# Install dependencies
bun install

# Start the development server
bun run dev

The app will be available at http://localhost:3000

Building for Production

bun run build
bun run start

Linting

bun run lint

Database

The project uses Prisma with SQLite. The database schema is in prisma/schema.prisma. To push schema changes:

bun run db:push

Project Structure

medisync/
|
|-- src/
|   |-- app/
|   |   |-- api/
|   |   |   |-- clinical-reasoning/route.ts   # Streaming LLM differentials + narrative
|   |   |   |-- chat/route.ts                 # Streaming LLM chat with RAG
|   |   |   |-- soap/route.ts                 # Streaming LLM SOAP note generation
|   |   |   |-- route.ts                      # Health check
|   |   |-- globals.css                       # TypeUI Tetris design system + dock + glassmorphism
|   |   |-- layout.tsx                        # Root layout with Bangers + JetBrains Mono fonts
|   |   |-- page.tsx                          # Main page (onboarding gate + zone router)
|   |
|   |-- components/
|   |   |-- medisync/
|   |   |   |-- app-shell.tsx                 # Sidebar + header + floating dock layout
|   |   |   |-- onboarding.tsx                # Full screen patient intake form
|   |   |   |-- patient-sidebar.tsx           # (Legacy, now in app-shell)
|   |   |   |-- store.ts                      # Zustand store (multi patient, localStorage, API calls)
|   |   |   |-- shared.ts                     # Triage colors, severity colors, helpers
|   |   |   |-- vitals-radar.tsx              # Hexagonal SVG radar component
|   |   |   |-- theme-provider.tsx            # (Legacy, light mode only)
|   |   |   |-- views/
|   |   |       |-- overview.tsx              # Clinical snapshot with vitals radar
|   |   |       |-- patient.tsx               # Clinical intake form
|   |   |       |-- reasoning.tsx             # Differentials + evidence + narrative
|   |   |       |-- tools.tsx                 # Chat + SOAP + What If + Knowledge tabs
|   |   |       |-- (14 legacy view files, not imported)
|   |   |-- ui/                               # shadcn/ui components
|   |
|   |-- lib/
|   |   |-- ai/
|   |   |   |-- clinical-llm.ts              # Direct API integration, JSON repair, streaming
|   |   |-- medical/
|   |   |   |-- types.ts                     # TypeScript types for all clinical data
|   |   |   |-- engine.ts                    # Deterministic clinical rules
|   |   |   |-- knowledge.ts                 # ICD 10, drugs, guidelines, knowledge graph
|   |   |-- utils.ts                         # cn() class merge utility
|   |   |-- db.ts                            # Prisma client
|   |
|   |-- hooks/
|       |-- use-mobile.ts
|       |-- use-toast.ts
|
|-- prisma/
|   |-- schema.prisma                        # Database schema
|
|-- public/
|   |-- logo.svg
|   |-- robots.txt
|
|-- package.json
|-- tailwind.config.ts
|-- tsconfig.json
|-- next.config.ts
|-- eslint.config.mjs
|-- Caddyfile                                # Gateway config

API Reference

POST /api/clinical-reasoning

Request:

{
  "patient": {
    "id": "PT-001",
    "name": "Robert Chen",
    "age": 67,
    "gender": "male",
    "symptoms": ["Chest pain", "Sweating"],
    "vitals": { "heartRate": 104, "systolicBP": 142, ... },
    "labs": { "troponin": 0.84, ... },
    ...
  }
}

Response: Server Sent Events stream

Heartbeat lines: : heartbeat
Final data: data: { "analysis": ClinicalAnalysis, "llmPowered": true, ... }

POST /api/chat

Request:

{
  "question": "What supports the leading diagnosis?",
  "patient": PatientInput,
  "analysis": ClinicalAnalysis,
  "history": [{ "role": "user", "content": "..." }, ...]
}

Response: Server Sent Events stream with final data containing content, reasoning, and citations.

POST /api/soap

Request:

{
  "patient": PatientInput,
  "analysis": ClinicalAnalysis
}

Response: Server Sent Events stream with final data containing the SOAP note JSON.

Design System

Color Tokens

Token	Hex	Usage
Background	#DFE7FF	Page surface (periwinkle)
Foreground	#1C398E	Body text (deep blue)
Primary	#1C202B	Borders, headings, structure (navy)
Secondary	#7107E7	Accents, active states (purple)
Success	#16A34A	Normal vitals, low risk
Warning	#D97706	Abnormal vitals, moderate risk
Danger	#DC2626	Critical alerts, high risk
Card	rgba(255,255,255,0.75)	Frosted glass cards
Muted	#5A6BB8	Secondary text, labels

Typography

Display: Bangers (headlines, stat numbers, section numbers)
Body: JetBrains Mono (all body text, data values, labels)
Label: JetBrains Mono uppercase 0.6875rem with 0.1em letter spacing

Triage Colors

Level	Color	Meaning
Red	#DC2626	Resuscitation (immediate)
Orange	#D97706	Emergent (1 hour)
Yellow	#D97706	Urgent (24 hours)
Green	#16A34A	Less urgent (3 days)
Blue	#7107E7	Non urgent (1 week)

Mobile Responsiveness

The app is fully responsive from 375px (iPhone) to 1280px+ (desktop):

Below 768px (Mobile):

Sidebar hidden, hamburger menu opens slide in drawer with overlay
Floating dock shrinks to icons only (no labels)
Header compacts (hides labels, shows status dot only)
Onboarding form uses 2 column grid, smaller text, full width submit
Overview stacks vertically (patient info above vitals radar)
Quick actions go 2x2 instead of 1x4
Vitals radar scales to fit container (max 280px)

768px and up (Desktop):

Sidebar always visible (240px fixed)
Full floating dock with labels
Full header with patient name and ID
Overview uses 2 column layout (patient info beside vitals radar)
Quick actions in 1x4 grid

Security and Compliance Notes

MediSync is a clinical decision support tool. It is NOT a medical device and must NOT be used for actual patient care without:

FDA clearance as Software as a Medical Device (SaMD)
Clinical validation against gold standard diagnoses
HIPAA compliance for patient data handling (currently uses localStorage only)
Audit logging for all AI generated recommendations
RBAC (role based access control) for doctors, nurses, admins
Integration with hospital EHR systems via HL7/FHIR

The LLM (GLM 4) is called server side only. Patient data is never exposed to the client beyond what is displayed. The LLM API calls are handled exclusively in API routes, never in client components.

What Would Need to Change for Production

Replace localStorage with a HIPAA compliant database (PostgreSQL with encryption)
Add NextAuth.js authentication with JWT and OAuth
Replace the 6 drug database with Lexicomp or Micromedex (50,000+ interactions)
Replace the 10 disease database with ICD 10/SNOMED CT (10,000+ conditions)
Replace heuristic risk scores with validated implementations
Replace the multi model comparison with real trained ML model endpoints
Add audit logs for every AI prediction and clinician action
Add rate limiting and API key management
Run clinical validation studies

Limitations

What is Real

GLM 4 LLM generates all differentials, narrative, chat responses, and SOAP notes
Clinical safety rules (vitals, drug interactions, triage, risk scores) use real medical heuristics
Medical knowledge base contains real ICD 10 codes, real drugs, real guideline excerpts
Streaming API with heartbeats prevents gateway timeouts
JSON repair handles common LLM output errors
Multi patient support with localStorage persistence
Mobile responsive from 375px to 1280px+

What is Heuristic (Not Real ML Models)

Multi model comparison: Shows 7 ML model architectures (XGBoost, CatBoost, LightGBM, etc.) with heuristic AUC/F1 scores. This is an explainability visualization, not real trained models. Would need actual model endpoints in production.
SHAP feature contributions: Uses hardcoded weight values based on clinical rules. These are heuristic SHAP style explanations, not real SHAP values from a trained model.
Differential diagnosis engine: The deterministic fallback uses a homemade Bayesian style symptom matcher with heuristic weights, not a validated diagnostic tool. The LLM is actually more clinically sound than this engine.
Synthetic patient generator: Generates 5 hardcoded patient scenarios. The "Add Patient" button creates a blank template, not a synthetic patient.

What is Not Clinically Validated

The rule based engine is a reasonable educational and demo tool that demonstrates the shape of clinical decision support. It is NOT clinically validated and must NOT be used for actual patient care. The probability scores are not calibrated. The drug database has only 6 drugs. The disease database has only 10 diseases. Real CDSS tools (Isabel, DXplain, UpToDate) use expert curated databases or ML models trained on millions of cases.

License

This project is for demonstration and educational purposes. All clinical guidelines referenced are publicly available from their respective organizations (ACC/AHA, SSC, AHA/ASA, ATS/IDSA, ADA, ESC, KDIGO, AHA).

MediSync is a decision support tool, not a replacement for clinical judgment. Final diagnosis and treatment decisions remain the responsibility of the treating clinician.

Built with Next.js 16, TypeScript, Tailwind CSS 4, shadcn/ui, Framer Motion, Zustand, and direct GLM 4 API integration.

Pipeline Override & Versioning (July 2026)

Day 6 on the forge: It quietly lied. Overrode preprocessing ? 24 models trained on ghost data. SHAP looked normal. Fixed with per-step version tags.

Pipeline Improvements

Added full step versioning system
Fixed stale data bug affecting 24 models
SHAP, Radar, and SOAP now update correctly on overrides

Name		Name	Last commit message	Last commit date
Latest commit History 45 Commits
db		db
examples/websocket		examples/websocket
prisma		prisma
scripts		scripts
src		src
.gitignore		.gitignore
CHANGELOG.md		CHANGELOG.md
Caddyfile		Caddyfile
README.md		README.md
components.json		components.json
eslint.config.mjs		eslint.config.mjs
next.config.ts		next.config.ts
package-lock.json		package-lock.json
package.json		package.json
postcss.config.mjs		postcss.config.mjs
tailwind.config.ts		tailwind.config.ts
tsconfig.json		tsconfig.json

Folders and files

Latest commit

History

Repository files navigation

MediSync

AI Powered Intelligent Clinical Triage and Decision Support Platform

Table of Contents

What MediSync Does

Key Features

Real LLM Clinical Reasoning

Deterministic Clinical Safety Rules

Multi Patient Management

Vitals Radar Visualization

AI Chat Assistant

SOAP Note Generation

What If Simulator

Knowledge Graph

Technology Stack

Frontend

Backend

Data

Design

Architecture Overview

Real AI Integration

Three API Endpoints

JSON Repair

Probability Normalization

Clinical Safety Rules

Vital Sign Analysis

Drug Interaction Checking

Allergy Cross Reactivity

Triage Scoring

Risk Scores

Admission Prediction

Medical Knowledge Base

Disease Database

Drug Database

Clinical Decision Rules

RAG Knowledge Corpus

Knowledge Graph

ICD 10 and CPT Codes

User Interface

Design System

Navigation

Layout

Onboarding Flow

Zones

Patient Workflow

Getting Started

Prerequisites

API Key Configuration

Installation

Building for Production

Linting

Database

Project Structure

API Reference

POST /api/clinical-reasoning

POST /api/chat

POST /api/soap

Design System

Color Tokens

Typography

Triage Colors

Mobile Responsiveness

Security and Compliance Notes

What Would Need to Change for Production

Limitations

What is Real

What is Heuristic (Not Real ML Models)

What is Not Clinically Validated

License

Pipeline Override & Versioning (July 2026)

Pipeline Improvements

About

Resources

Uh oh!

Stars

Watchers

Forks