Medx AI

HEALTHCARE SYSTEM FOR BREAST CANCER DIAGNOSIS WITH EXPLAINABLE AI AND PERSONALISED LLM RECOMMENDATIONS

Bakary Gibba · Md Rashidul Islam · Faisal Rafat Faisal Khrisan
Bachelor of Computer Science (Honours), Albukhary International University

Important

Medx AI is a research prototype for clinical decision support. It does not provide medical diagnoses or replace pathologists, oncologists, confirmatory testing, or institutional clinical protocols.

The Project

Medx AI explores how an AI-assisted HER2 workflow can move beyond a standalone prediction model. The platform combines four-class histopathology classification, uncertainty-aware input checks, visual explanation, contextual LLM-generated recommendation drafts, and explicit clinician approval in one role-based web system.

Reported model performance	Evaluation scope	Clinical roles	Public repository scope
~96% weighted accuracy	154 tests passed	Doctor · Nurse · Admin	Case study only
4 HER2 classes	4.5/5 average UAT	Separated responsibilities	No source or dataset

Why It Matters

HER2 assessment influences breast cancer treatment planning, but ambiguous staining, observer variability, resource constraints, and opaque AI predictions can limit confidence. Medx AI investigates a more responsible pattern:

flowchart LR
  A[Histopathology image] --> B[Quality & distribution checks]
  B --> C[HER2 ensemble prediction]
  C --> D[Confidence & uncertainty]
  C --> E[Grad-CAM explanation]
  D --> F[Context-aware LLM draft]
  E --> F
  F --> G{Clinician review}
  G -->|Accept| H[Traceable report]
  G -->|Reject| I[Revise or investigate]

Loading

The key design decision is the review boundary: model and LLM outputs remain assistive drafts, while clinicians retain final responsibility.

Research Highlights

Area	Portfolio evidence
Classification	Soft-voting ensemble across Random Forest, Logistic Regression, and SVM for HER2 0, 1+, 2+, and 3+
Explainability	Grad-CAM overlays place visual evidence beside predictions instead of hiding interpretation in a separate research notebook
Safety	Blank-slide, blur, stain-profile, distribution-shift, and entropy checks can reject or defer uncertain inputs
Generative AI	Structured LLM drafts incorporate prediction context, optional notes, and extracted patient-history text
Human oversight	Recommendations follow draft, review, acceptance, rejection, and reporting states
System design	Role-based dashboards, audit history, object storage, model versioning, and traceable clinical records

Selected Experience

Prediction & XAI	Clinician review

Role-based operations	Case lifecycle visibility

Browse all sanitized screenshots →

Reported Evaluation

HER2 class	Precision	Recall	F1	Average precision
0	0.99	0.93	0.96	0.994
1+	0.87	0.96	0.92	0.950
2+	0.87	0.95	0.91	0.953
3+	1.00	0.98	0.99	0.999

Note

These are project-report results from a controlled academic evaluation, not evidence of prospective clinical validity. External validation across institutions, scanners, demographics, and staining protocols remains necessary.

My Contribution

Bakary Gibba's assigned responsibility was model development and explainable AI. The work included the image-processing and classification pipeline, ensemble design, class-imbalance strategy, prediction confidence and safety behavior, Grad-CAM integration, model evaluation, and contribution to documentation and system testing.

The project was collaborative: Md Rashidul Islam led frontend development, Faisal Rafat Faisal Khrisan led backend development, and all team members contributed to documentation and testing. See My Contributions for the precise ownership boundary.

Technology Landscape

Layer	Technologies and patterns
AI and imaging	Python, scikit-learn, PyTorch, OpenCV, OpenSlide, Grad-CAM, entropy analysis
Generative AI	Groq-hosted Llama, Gemini fallback, structured prompting, OCR-assisted context
Application	React, TypeScript, Vite, Tailwind CSS, Django, Django REST Framework
Data and storage	PostgreSQL-oriented relational design, Django ORM, MinIO object storage
Security and governance	JWT, RBAC, 2FA-ready flows, audit logging, soft deletion, model registry

Case Study Navigation

Research	Engineering	Responsible AI
Executive Summary	System Architecture	Safety & Ethics
Clinical Problem	Data & System Design	Challenges
Research Methodology	Clinical Workflows	Future Work
AI Model	Evaluation	Appendix
Explainable AI	My Contributions	Diagram Library
LLM Recommendations	Screenshot Gallery	Documentation Index

Lessons Learned

• Accuracy alone is insufficient in high-stakes AI; uncertainty, explanation, traceability, and escalation must shape the workflow.
• Clinician review is a system state and accountability boundary, not a disclaimer added after generation.
• Data-efficient models can be valuable when compute and labeled medical data are limited, but external validity matters more than a single benchmark.
• LLMs are better positioned as structured communication and decision-support layers than autonomous clinical authorities.

---

Public academic case study · January 2026

GitHub Profile · Portfolio · LinkedIn

This repository intentionally excludes source code, model weights, credentials, the private project report, patient records, and the underlying research dataset.