PrismDX — AI Diagnostic Bias Detection Tool

Google Solution Challenge 2026 — SDG 3: Good Health and Well-Being

A full-stack medical AI tool that analyses diagnostic images and flags potential demographic bias in AI confidence scores, cross-referenced against peer-reviewed Fitzpatrick-stratified baselines.

🔗 Live Demo: https://prismdx-2026.web.app

---

What It Does

Medical AI models have documented accuracy disparities across demographic groups — particularly across Fitzpatrick skin types (up to 23% lower accuracy for Types V–VI). PrismDX makes this visible at the point of care by:

• Running AI image analysis via Gemini 2.5 Flash
• Cross-referencing AI confidence against published bias baselines (Daneshjou et al., Nature Medicine 2024; Seyyed-Kalantari et al., 2021)
• Flagging scans where confidence falls below the expected threshold for that demographic — or where confidence is suspiciously high on historically underserved skin types
• Tracking scan history with search, filter, and PDF export
• Providing clinicians with a one-click human review request tied to the audit trail

---

Tech Stack

Layer	Technology
Frontend	Next.js 15, React 19, Tailwind CSS, shadcn/ui
Backend	FastAPI, Python 3.13
AI Model	Google Gemini 2.5 Flash (gemini-2.5-flash) with gemini-2.0-flash fallback
Hosting	Firebase Hosting
Backend Hosting	Railway
Charts	Recharts
State	localStorage + sessionStorage

---

Features

• New Scan — Upload a medical image with patient demographics (age, gender, Fitzpatrick scale, body localization)
• Scan Types — Dermoscopy, Skin Lesion, Chest X-ray, Mammography, CT Scan, MRI
• Image Validation — Gemini validates uploads are genuine medical images before analysis
• Bias Evaluation — Confidence vs Fitzpatrick baseline comparison with risk level (low / moderate / high)
• Bias Flag — Triggers when confidence is below threshold for the demographic, or unusually high on high-risk Fitzpatrick types (V/VI)
• Dashboard — Live stats (total scans, bias flags, avg confidence, clear rate) and confidence-by-Fitzpatrick chart
• History — Search, filter by bias flag / scan type / risk level, sort, delete, export to PDF
• Human Review — One-time flag to request clinical verification, persisted in history
• PDF Export — Full diagnostic report including bias evaluation and recommendations

---

Project Structure

PrismDX/
├── Backend/
│   ├── main.py           # FastAPI app + /scan endpoint + bias logic
│   ├── requirements.txt  # Python dependencies
│   └── .env              # GEMINI_API_KEY (not committed)
└── Frontend/
    ├── firebase.json     # Firebase hosting config (cleanUrls enabled)
    ├── app/
    │   ├── layout.tsx
    │   └── (app)/
    │       ├── page.tsx            # Dashboard
    │       ├── layout.tsx          # Shared sidebar layout
    │       ├── scan/page.tsx       # New scan form
    │       ├── results/page.tsx    # Diagnosis + bias results
    │       ├── history/page.tsx    # Scan history
    │       └── methodology/page.tsx
    ├── components/
    ├── lib/
    │   ├── api.ts          # API types + fetch helper
    │   └── pdf-export.ts   # PDF report generator
    └── .env.local          # API URL (not committed)

---

Setup & Running

Prerequisites

• Python 3.10+
• Node.js 18+
• Gemini API key — free at aistudio.google.com

Backend

cd Backend
pip install -r requirements.txt

Create a .env file in the Backend/ folder:

GEMINI_API_KEY=your_gemini_api_key_here

Run the backend:

uvicorn main:app --reload --port 8000

Frontend

cd Frontend
npm install

Create a .env.local file in the Frontend/ folder:

NEXT_PUBLIC_API_URL=http://localhost:8000

Run the frontend:

npm run dev

Open http://localhost:3000

Deploy to Firebase

cd Frontend
npm run build
firebase deploy --only hosting

Note: Ensure firebase.json is in the root of the project with "public": "Frontend/out" and "cleanUrls": true.

---

Bias Detection Logic

Bias flags are raised in two scenarios:

1. Underconfidence — AI confidence falls below the published threshold for the patient's Fitzpatrick type (classic bias pattern)
2. Overconfidence on high-risk types — AI confidence is suspiciously high (>10% above baseline) for Fitzpatrick Types V or VI, which are historically underrepresented in training data

Baselines Reference

Derived from peer-reviewed research:

Fitzpatrick Type	Skin Description	Dermoscopy Baseline	Skin Lesion Baseline
Type I	Pale white	92%	94%
Type II	White	91%	93%
Type III	Light brown	88%	90%
Type IV	Moderate brown	83%	86%
Type V	Dark brown	74% ⚠️	78% ⚠️
Type VI	Deeply pigmented	67% ⚠️	71% ⚠️

Sources:

• Daneshjou et al. — Disparities in dermatology AI performance on a diverse, curated clinical image set. Nature Medicine, 2024
• Seyyed-Kalantari et al. — Underdiagnosis bias of artificial intelligence algorithms applied to chest radiographs in under-served patient populations. Nature Medicine, 2021

---

API Endpoints

Method	Endpoint	Description
POST	/scan	Analyse image + return diagnosis and bias evaluation
GET	/health	Health check + active model info
GET	/baselines	Full bias baselines JSON for all scan types

POST /scan — Form Fields

Field	Type	Required	Description
image	file	Yes	Medical image (JPEG/PNG/WebP, max 50MB)
scan_type	string	Yes	skin-lesion, dermoscopy, chest-xray, mammography, ct-scan, mri
fitzpatrick	string	Yes	1 through 6
age	string	No	Patient age
gender	string	No	Patient gender
localization	string	No	Body location (relevant for skin scans)

Example Response

{
  "condition": "Melanoma (highly suspicious)",
  "finding_detected": true,
  "confidence": 90.0,
  "baseline_confidence": 88.0,
  "has_bias_flag": false,
  "bias_risk_level": "low",
  "bias_explanation": "Confidence within expected range for Type III.",
  "fitzpatrick_label": "Type III — Light brown, sometimes burns",
  "model_version": "PrismDX v3.0.0 / gemini-2.5-flash",
  "timestamp": "2026-04-28T11:39:58.214Z"
}

---

Disclaimer

PrismDX is a research prototype built for the Google Solution Challenge 2026. It is not a substitute for professional medical diagnosis. All results must be reviewed by a qualified clinician before informing any medical decision.