Machine learning / Data science / Academic research
Private logistic-regression research experiment exploring binary classification, ROC visualization and explicit input validation for tabular data.
Original code-generated category illustration; not a product screenshot.
Private codebase; this page shares a general technical summary. Source files, real records and organization-specific information remain private.
A logistic-regression research workflow for tabular data in a sleep-apnea context. It includes input loading, complete-row filtering, a stratified train/test split and evaluation plots. The project distinguishes software behavior from scientific validity and makes no claim of clinical success or performance measured on real patient data.
Incorrect target labels, nonnumeric fields and accidental use of the target as an input can invalidate small research experiments. Successful execution is also not evidence of clinical benefit. The project explores a basic classification workflow with explicit input contracts and bounded software checks.
Research data preparation and linkage of anthropometric and clinical variables; Python/pandas processing, classification and model evaluation.
CSV or spreadsheet inputs are processed against numeric-feature and binary-target requirements. After incomplete rows are removed, a stratified holdout split supports logistic-regression evaluation. Classification summaries, ROC plots and coefficient inspection are research outputs. Review-branch checks reject missing columns, invalid labels and direct selection of the target as an input feature.
My research contribution included working with anthropometric measurements extracted from Tanita reports alongside clinical variables, AHI/OSAS classes and questionnaire results. I used Python and pandas for cleaning and linkage. This project describes the tabular classification experiment associated with that broader preparation work; it is a separate software component from the report-extraction application.
The workflow explicitly identifies feature columns and the target before classification. Non-numeric inputs, missing columns and labels outside the binary classes make results harder to interpret. Rejecting the target column in the feature list provides a basic safeguard against direct target leakage. Other indirect leakage and related records from the same participant require additional research-design controls.
A stratified training/test split precedes logistic regression, a classification summary, an ROC plot and coefficient inspection. Coefficients offer a starting point for examining how the model uses variables; they do not establish causality. Synthetic software tests check execution and rejection of invalid inputs. External validation, calibration and clinical-use evaluation remain outside this scope.
Load authorized tabular research inputs
Check numeric features and binary-target requirements
Apply a stratified holdout split and fit logistic regression
Inspect classification summaries, ROC plots and coefficients
Conceptual flow derived from project descriptions; not an application screenshot.
Private document-processing prototype exploring local PDF-to-table extraction through batch processing and a desktop interface, with explicit privacy boundaries.
Private research toolkit exploring EEG, signal features and classical machine learning through desktop, offline and local web components.