Respiratory disease classification

Respiratory diseases can share symptoms, making classification a useful problem for studying model evaluation. The work uses a public Kaggle dataset and organizes exploratory analysis, feature selection, and algorithm comparison in a notebook within an academic setting.

I developed the study with Caio Moura Portela de Sousa, contributing to data preparation, modeling, evaluation, and interpretation. The notebook brings together decision trees, Random Forest, logistic regression, and an MLP, with metrics and visualizations for comparing their behavior.

1. The dataset is filtered to the three selected diseases and symptoms with no occurrences are removed.
2. A stratified split reserves 30% for testing; feature count is selected through a feature-selection pipeline and cross-validation on training data.
3. Four models are trained and evaluated using confusion matrices, precision, recall, F1, and balanced accuracy.
4. The analysis compares errors, feature importance, and train-test differences, exporting a metrics table.

Evaluating a model requires discussing data origins, feature selection, and error types alongside the final metric. The study showed how a linear approach can be competitive and how interpretation depends on the experimental setting.

• An academic study without clinical validation; the binary-symptom dataset does not support use as a diagnostic tool.
• The repository does not include the dataset or pinned dependency versions. The metrics presented are saved results, without guaranteed reproduction in another environment.

João Matheus de Figueirêdo Tavares and Caio Moura Portela de Sousa · Information Systems, Unifacisa.