Thyroid Disorder Prediction using Machine Learning

Thyroid disorders — hypothyroidism and hyperthyroidism — affect hundreds of millions of people, and diagnosis starts with one routine blood panel: TSH, T3 and T4. Because the hormones regulate each other through the body's feedback loop, the pattern matters more than any single number: suppressed TSH with elevated T4 points one way, raised TSH with low T4 the other. Students usually learn this as a memorized table, never as data. This project turns it into machine learning on the well-known UCI Thyroid Disease dataset (~7,200 records, 29 attributes including T3, T4 and TSH). A web demo makes the model tangible: entered lab values return the predicted disorder class with its full probability distribution. Scope is stated honestly — an educational prototype for learning, never a certified diagnostic device.

1. The thyroid-disease data (~7,200 records, 29 attributes) is loaded; T3/T4/TSH columns are cleaned, missing values imputed with medians, and categorical fields one-hot encoded.
2. The data is split 80/20 train/test, stratified so the class balance (hyperthyroid, hypothyroid, normal) is preserved in both halves.
3. Four classifiers — Logistic Regression, Random Forest, XGBoost and SVM — are compared with stratified 5-fold cross-validation on macro-F1.
4. The winning model is retrained on the full training set, then evaluated once on the held-out test split: confusion matrix plus per-class precision, recall and F1.
5. Permutation importance ranks the driving features; TSH and T4 lead, which matches how the thyroid feedback loop works in practice.
6. The trained model is exported (.pkl) with its preprocessing pipeline and wired into the demo: entered lab values pass through the identical preprocessing and return class probabilities in real time.

Tech stack:

• Python 3, scikit-learn
• XGBoost
• Pandas, NumPy
• Jupyter Notebook (training & evaluation)
• Matplotlib, Seaborn (visualization)
• HTML · CSS · JavaScript (screening demo)
• UCI Thyroid Disease dataset (Quinlan et al.)

Dataset & model details

• Dataset: UCI Machine Learning Repository thyroid-disease dataset (Quinlan et al.) — ~7,200 patient records with 29 attributes covering T3, T4 and TSH measurements plus demographic and clinical fields.
• Task: 3-class classification; input = encoded lab-value feature vector, output = probability distribution over hyperthyroidism, hypothyroidism and normal.
• Model: XGBoost classifier (n_estimators 300, max_depth 6, learning rate 0.05), selected after 5-fold cross-validation comparison against Logistic Regression, Random Forest and SVM.
• Metrics: Test macro-F1 ~0.94 is the design target for the built-to-order training run; per-class precision/recall/F1 and the confusion matrix are measured on the held-out test split. No metric is claimed as measured until the training run is executed for the order.

• [Lab-value screening demo] Enter TSH, T3, T4, age and sex and get the predicted disorder class — hyperthyroidism, hypothyroidism or normal — with a confidence meter and per-class probability bars.
• [Full training notebook] Data loading, missing-value handling, exploratory analysis, encoding, model training and evaluation in one reproducible Jupyter notebook.
• [Model comparison] Logistic Regression, Random Forest, XGBoost and SVM compared with stratified 5-fold cross-validation; selection on macro-F1.
• [Confusion matrix & per-class report] Precision, recall and F1 on the held-out test split, with the cost of misclassification discussed in a medical-screening context.
• [Permutation feature importance] Model-agnostic ranking showing TSH and T4 carrying the most signal — matching clinical intuition about the thyroid feedback loop.
• [Batch screening mode] Score an intake queue of patient records at once and flag those above the decision threshold.
• [Calibrated decision threshold] The operating point is chosen on the validation split with its trade-offs documented.
• [Educational medical disclaimer] Every demo screen states the prototype is for learning and never replaces clinical diagnosis.

• Complete training & evaluation Jupyter notebook
• Trained classifier (.pkl) with preprocessing pipeline
• Lab-value screening web demo with probability bars
• Confusion matrix, per-class metrics and feature-importance plots
• Project report PDF (thyroid physiology background, methodology, results)
• PPT presentation for final review
• Viva Q&A preparation document (classification metrics, cross-validation, class imbalance, overfitting)

• The model is an educational prototype, not a certified medical device; its predictions do not replace clinical diagnosis.
• Trained on a public research dataset — real clinic data differs in demographics, assay methods and reference ranges.
• 0.94 macro-F1 is a design target, stated honestly — the report documents the actual achieved figure after the order's training run.
• The interactive demo simulates the trained model in JavaScript; the full trained Python pipeline ships with the order.
• Borderline subclinical cases (TSH only slightly outside range) are the most likely to be misclassified; per-class recall is quantified in the report.
• The feature set is limited to 29 dataset attributes — it cannot see symptoms, medication or history the way a doctor would.

Suitable for B.E./B.Tech final-year projects in Computer Science, IT and AI & Machine Learning.