Predicting disease risk from clinical data
Comparative machine learning frameworks for cardiovascular, metabolic, oncological and neurodevelopmental disease prediction. Independent projects, Jan to Mar 2026.
- 4conditions, 4 repositories
- 768diabetes records, 8 features
- 500 / 268non-diabetic / diabetic
- +5.1 ptstest accuracy gain, trial 1 to 3
- 85.7%heart ANN test accuracy
Live app and notebook
The interactive version of these projects: a multipage Streamlit app (Home, Diabetes, Breast cancer, Heart disease, Autism, Compare) plus the Colab notebook. This static page documents the analysis below; the app lets you rerun it with your own settings.
Live Streamlit app Diabetes Breast cancer Heart disease Autism Compare Open in Colab Notebook (.ipynb) GitHub repo
| What | Link |
|---|---|
| Live app (Streamlit Cloud) | ml-healthcare-ai.streamlit.app |
| This website (GitHub Pages) | harshini-suresha.github.io/ml-healthcare-app |
| Run notebook (Colab) | Open in Colab |
| Notebook file (.ipynb) | Download raw .ipynb · nbviewer preview |
| Repo | github.com/Harshini-Suresha/ml-healthcare-app |
Run locally: pip install -r requirements.txt then streamlit run app.py. Multipage app: Home (app.py), shared code (ui_common.py, ml_core.py), pages in pages/ including the A-vs-B Compare diff view.
Projects at a glance
Each project follows the same path: exploratory data analysis, feature engineering, preprocessing, model development and quantitative evaluation.
- Explore
- Engineer features
- Preprocess
- Train
- Tune
- Evaluate
- Save model
Workflow map
Highlighted stage is the model. Stages follow each repository's description.
| Project | Area | Approach | Data | Code |
|---|---|---|---|---|
| Heart disease | Cardiovascular | Artificial neural network (13-8-14-1), nonlinear classification | 303 patient records, 13 attributes | Heart-disease |
| Diabetes | Metabolic | K-Nearest Neighbors, three-trial comparative study | Kaggle diabetes dataset, 768 records | Diabetes |
| Breast cancer | Oncology | KNN with cross-validated K versus tuned Logistic Regression | Clinical diagnostic features, malignant vs benign | Breast-Cancer-Prediction |
| Autism (ASD) | Neurodevelopmental | Logistic Regression, XGBoost and SVM compared on a rebalanced dataset | 800 screening records, 22 columns | Autism-prediction |
Diabetes: a three-trial KNN study
Same dataset every time. Each trial changed preprocessing, tuning or validation, so the accuracy change can be traced to a cause.
The dataset
768 patient records, 8 diagnostic features and a binary outcome. Zeros in five columns are not real measurements; they stand in for missing values.
| Feature | Meaning |
|---|---|
| Pregnancies | Number of times pregnant |
| Glucose * | Plasma glucose concentration |
| BloodPressure * | Diastolic blood pressure (mm Hg) |
| SkinThickness * | Triceps skin fold (mm) |
| Insulin * | 2-hour serum insulin (mu U/ml) |
| BMI * | Body mass index (kg/m²) |
| DiabetesPedigreeFunction | Genetic likelihood of diabetes |
| Age | Age in years |
| Outcome | 1 diabetic, 0 non-diabetic |
* a zero in this column means the value is missing.
Results by trial
Exploring the data: real statistics from the 768 records
Computed from the public copy of the dataset (768 rows, 9 columns). Pick a feature to update the histogram and box plot. Zeros in the five missing-as-zero columns are excluded from them.
Correlation matrix
Pearson r. Teal is positive, red is negative.
Correlation with Outcome
Zeros that mean missing
Share of 768 records. Insulin is missing for nearly half of them.
Explore the real records: scatter plot and live KNN
All 768 records. Pick two features, then click the plot to place a new patient. KNN votes among the k closest records. Records where a missing-as-zero feature is 0 are left out. Switch to standardized distances and watch the neighbours change: with raw units the feature with the larger numbers dominates.
● non-diabetic ● diabetic ◯ new patient
Raw versus standardized distances, measured on the real data
Leave-one-out KNN accuracy on all 768 records using all 8 features, shown below and recomputed live by the button. Each record is classified from the others. This uses no cleaning or train/test split, so it will not match the trial results below. Standardizing helps most at small k (for example k=3 and k=5); at larger k the two settle at about the same accuracy.
The trial journey, from your reported results
Test accuracy (axis starts at 60). Most of the gain came from scaling and outlier filtering in trial 2.
Train minus test accuracy, in points. A wider gap means more memorization.
Best k per trial against train (grey) and test (teal) accuracy. Smaller k fits the training data tighter.
What changed in each trial
The train-test gap grew from 18.4 to 20.2 to 24.8 points while test accuracy rose. Trial 3 uses k=1, which memorizes the training set (100% train accuracy), so the 75.2% test figure is the one that counts.
Pipeline code
Breast cancer: KNN versus a tuned Logistic Regression
Predicts malignant versus benign from clinical diagnostic features. The focus is on how preprocessing and model configuration change performance, with both models chosen by cross-validation instead of by hand.
Workflow
Preprocessing
- Missing values found and replaced with the column mean
- Features and target separated
- StandardScaler so every feature contributes comparably
- Separate hold-out set for testing
K-Nearest Neighbors
- Euclidean distance, so scaling matters
- K chosen by cross-validation, not guessed
- Small K fits tightly, large K smooths: the bias-variance trade-off
Logistic Regression
- Linear probabilistic classifier
- GridSearchCV tries every parameter combination
- RepeatedStratifiedKFold keeps class proportions in each fold and repeats the split
Evaluation
- Accuracy
- Precision
- Recall
- F1-score
- Confusion matrix
- Classification report
Educational implementation, not for clinical diagnosis.
Heart disease: an artificial neural network
A feedforward network that classifies heart disease from 13 clinical attributes. 303 patient records, 54.5% with the disease, no missing values.
Architecture
Two ReLU hidden layers and a sigmoid output. Adam optimizer, binary cross-entropy, batch size 8, 100 epochs. Inputs standardized, 70/30 split.
Training over 100 epochs
Training accuracy climbs from 49.7% to 92.1% and loss falls from 0.693 to 0.259.
Test results: 91 held-out patients
| Accuracy | 85.7% | Precision | 82.7% |
|---|---|---|---|
| Recall | 91.5% | F1-score | 86.9% |
Recall is the strongest score: 43 of 47 patients with the disease were caught, 4 missed.
Beyond the network
- Random forest feature ranking puts oldpeak, cp, thalach, ca and thal on top.
- GridSearchCV on a random forest picked 300 trees, max depth 10, min split 2.
- 5-fold cross-validation of logistic regression averages 82.5% accuracy.
- age
- sex
- cp
- trestbps
- chol
- fbs
- restecg
- thalach
- exang
- oldpeak
- slope
- ca
- thal
Training milestones
Beyond accuracy: six scores from the confusion matrix
Specificity (79.5%) is the weakest score: 9 healthy patients were flagged. The model leans toward catching disease.
The patients: 303 records
Box: middle half of patients, line: median, whiskers: min to max, each on its own scale.
Figures from the notebook
Autism spectrum disorder: screening-based classification
800 records with ten screening-question scores plus demographic fields. About 20% are ASD-positive (161 of 800), so the data is heavily imbalanced and the evaluation accounts for it.
Cleaning
- Unknown ethnicity coded "?" (203 records) merged with "others"
- yes/no turned into 1/0
- 2 extreme result values removed (800 to 798)
- ID, age_desc, used_app_before and austim dropped from the features
Engineered features
- sum_score: total of the ten screening scores
- ageGroup: toddler, kid, teenager, young, senior
- ind: autism, app-use and jaundice flags combined
- Log transform of age to remove skew
Training setup
- 80/20 train-validation split
- Random oversampling of the minority class on the training set only
- StandardScaler fitted on the training data
- Models: Logistic Regression, XGBoost, SVM (RBF)
Model comparison
ROC-AUC computed from predicted labels. XGBoost scores 1.00 on training data but 0.75 on validation, a clear sign of overfitting. SVM generalizes best (0.80); Logistic Regression has the smallest gap.
Logistic Regression on 160 validation records
| Accuracy | 80.0% | ASD recall | 75.0% |
|---|---|---|---|
| ASD precision | 54.0% | Missed | 9 of 36 |
From 800 records to the training set
● no ASD ● ASD. Oversampling copies minority rows in the training set only, so the validation set stays untouched.
How often each screening answer is 1
Share of the 800 records scoring 1 on each question.
Who is in the data
Ethnicity, before merging "?" and "others"
Who filled in the form
Overfitting by model
Training minus validation ROC-AUC. Smaller is better.
Logistic Regression, six scores
Figures from the notebook
Findings and methods across the four projects
Diabetes. Scaling and outlier filtering lifted test accuracy 4.5 points (70.1% to 74.6%). Everything added in trial 3 gained 0.6 more, while the train-test gap widened to 24.8 points.
Heart. 85.7% accuracy and 91.5% recall on 91 held-out patients, with 4 cases missed. Final-epoch training accuracy of 92.1% is 6.3 points above test.
Autism. Only 20% of records are ASD-positive. XGBoost scores 1.00 ROC-AUC on training data but 0.75 on validation; SVM generalizes best at 0.80.
Breast cancer. K and the Logistic Regression settings are chosen with cross-validation and grid search instead of one lucky split.
Methods side by side
| Diabetes | Breast cancer | Heart | Autism | |
|---|---|---|---|---|
| Scaling | StandardScaler, MinMaxScaler | StandardScaler | StandardScaler | StandardScaler |
| Missing values | Inf / NaN rows dropped | Mean imputation | None missing | "?" recoded as Others |
| Outliers | Filtered, extremes removed | – | – | 2 extreme results removed |
| Class imbalance | Class balance checked | Stratified folds | SMOTE explored | Random oversampling |
| Feature engineering | – | – | – | Sum score, age group, log age |
| Models | KNN | KNN, Logistic Regression | ANN, plus RF, LR, SVM tried | LR, XGBoost, SVM |
| Tuning and validation | k scan 1 to 14, 70/30 split over seeds | CV for K, GridSearchCV, repeated stratified CV | GridSearchCV (RF), 5-fold CV (LR) | 80/20 hold-out |
| Reported scores | Accuracy, matrix, report | Accuracy, P, R, F1, matrix, report | Accuracy, P, R, F1, matrix | ROC-AUC, matrix |
How models are judged
| Metric | Formula | What it answers |
|---|---|---|
| Accuracy | (TP + TN) / all | How often the model is right overall. Misleading when classes are unbalanced, as in the 500 / 268 diabetes data. |
| Precision | TP / (TP + FP) | Of the patients flagged, how many truly have the condition. |
| Recall | TP / (TP + FN) | Of the patients who have it, how many were caught. The key screening metric. |
| F1-score | 2PR / (P + R) | One number that balances precision and recall. |
| Confusion matrix | TP FP / FN TN | Where the errors fall, split by type. |
| ROC-AUC | area under TPR vs FPR | Ranking quality across every decision threshold. |
Scores reported in the notebooks: accuracy, precision, recall and F1 (heart), ROC-AUC and a confusion matrix (autism), accuracy and a classification report (diabetes).
Contact
Questions about the projects, or want to collaborate?
All links
Every live URL for this project in one place.