Explainable Cancer Diagnosis ML

Overview

A governed ML evidence path spanning deterministic data splits, validation-only selection, locked evaluation, malignant-class explainability, checksum validation, strict API contracts, and a read-only frontend showcase.

Problem

ML portfolio results are difficult to trust when model selection, test evaluation, class semantics, and generated artifacts are not governed as one reproducible contract.

Solution

Built a deterministic 398/85/86 split, selected Logistic Regression using validation ROC-AUC only, evaluated the frozen model once on 86 governed test rows, and aligned malignant-class scores and SHAP explanations across the pipeline, FastAPI, and React.

Technical Decisions

• The split assignment is pinned by SHA-256 497e9350c039abd8f56c26e0fd3d6abf962bb8008fce5379f0b1790a9684df9c.
• The UI calls outputs uncalibrated malignant-class model scores, never confidence, clinical risk, or individual probability.

Outcome

The locked evaluation records ROC-AUC 0.9953703703703703, PR-AUC 0.99375, malignant sensitivity 0.96875, specificity 1.0, and confusion matrix [[31, 1], [0, 54]], with a read-only public showcase and explicit educational limits.

Key Features

• Validation ROC-AUC alone selects the model before the governed test result is inspected.
• Malignant-class score and SHAP semantics are explicit and tested across model, API, and frontend contracts.
• A schema-1 manifest validates 16 artifacts and fails closed on missing, stale, path-invalid, or checksum-invalid evidence.

Architecture

1. 01

   Governed dataset split

2. 02

   Train-only preprocessing

3. 03

   Validation-only selection

4. 04

   Locked evaluation

5. 05

   Malignant-class SHAP

6. 06

   Artifact manifest

7. 07

   FastAPI and read-only React

Tech Stack

• Python
• scikit-learn
• SHAP
• FastAPI
• React
• PyTorch
• Artifact governance

Verification

• Deterministic split: 398 train, 85 validation, 86 governed test rows
• Selected model: logistic_regression; evidence version: bbb5977c47501cd9a962
• Locked metrics: ROC-AUC 0.9953703703703703, PR-AUC 0.99375, sensitivity 0.96875, specificity 1.0
• Class order [malignant, benign]; confusion matrix [[31, 1], [0, 54]]

Security & Privacy

• The public Vercel surface is a read-only frontend showcase with no public live inference.
• FastAPI validates trusted repository-generated artifacts before loading code-bearing model files.

Limitations

• This is an educational and non-clinical portfolio system, not a medical device or diagnostic product.
• The small bundled dataset has no external, prospective, fairness, demographic, or clinical validation.
• Scores are uncalibrated and SHAP explains model behavior, not biology or causality.

Future Improvements

• Add approved external validation without changing the locked portfolio regression contract.
• Study calibration and subgroup behavior before considering any broader interpretation.