Back to Projects Projects / Predictive Analytics
Diagnostic Diabetes Predictor Preview
Predictive Analytics 2024 · Supervised Classification

Diagnostic Diabetes Predictor

A supervised classification benchmark evaluating logistic regression, support vector machines, and random forests on clinical metabolic markers for early diabetes risk identification.

Project Overview

Type-2 diabetes frequently remains undetected during early glycemic shifts, causing insidious microvascular and cardiovascular damage before a formal clinical diagnosis is rendered.

This diagnostic pipeline performs rigorous comparative analysis across standard supervised statistical learning models on metabolic indicators (fasting blood glucose, insulin, body mass index, age, and diabetes pedigree function). The project rigorously evaluates trade-offs between model interpretability and predictive power, delivering validated screening recommendations.

Key Features

Comparative Benchmark Suite

Benchmarked Logistic Regression, RBF-kernel SVM, Random Forests, and k-NN under uniform cross-validation folds.

Outlier & Zero-Value Treatment

Replaced biologically implausible zero-value records (e.g. 0 mm skinfold or 0 mg/dL blood pressure) with median imputation stratified by age bracket.

Threshold Tuning for Sensitivity

Optimized classification decision thresholds to minimize false-negative rates, prioritizing clinical patient safety in diagnostic screening.

Feature Importance Analysis

Quantified Gini impurity reduction and coefficient odds ratios to confirm glucose concentration and BMI as premier diagnostic indicators.

Technical Challenges Faced

Hidden Missing Data Encoded as Zeroes: Raw biomedical datasets frequently store null tests as numerical zeros. Detected and cleansed unnatural distributions in serum insulin and triceps skin fold thickness before model training.
Feature Scaling Disparities: Insulin and glucose numerical scales dwarfed pedigree coefficients. Standardized numerical attributes with z-score normalization to ensure gradient descent stability for SVM margins.
False Negative Penalization: Missing a diabetic patient carries higher clinical penalty than a false alarm. Applied asymmetric cost matrices during cost-sensitive classification to penalize false negatives threefold.

Improvements & Future Roadmap

Longitudinal HbA1c Tracking: Expanding feature schemas to ingest sequential 3-month HbA1c trajectory data for time-series risk forecasting.
API Deployment via FastAPI: Containerizing the top-performing estimator into a Dockerized REST microservice for integration with hospital electronic health record (EHR) platforms.