🩺 Classifying Medical Transcriptions by Specialty

About this simulation

Real clinical transcription datasets (like MTSamples) contain thousands of free-text notes labelled by medical specialty — cardiology, orthopedics, neurology, gastroenterology, radiology. A common lightweight NLP pipeline turns each note into a numeric vector with TF-IDF (term frequency × inverse document frequency, which down-weights common words and up-weights terms distinctive to a specialty), then trains a Logistic Regression classifier on those vectors.

This scene represents each transcription note as a point floating above a coloured floor tile. The floor is a 2D projection of the high-dimensional TF-IDF space, shaded by which specialty the current classifier would predict for any location — the decision surface. Well-separated, tightly clustered points with crisp boundaries mean the model classifies confidently; overlapping points near a boundary mean the note is ambiguous, which is exactly where a real classifier makes mistakes.

How it works

• TF-IDF vs raw frequency — TF-IDF suppresses generic words ("patient", "history") and keeps specialty-distinctive terms ("myocardial", "arthroscopy"), which tightens clusters. Raw term frequency keeps common words dominant, so clusters overlap more.
• Classifier — Logistic Regression draws straight (linear) boundaries between specialties. Naive Bayes assumes word independence and tends to carve boundaries that favour classes with more distinctive vocabulary, shifting territory unevenly.
• Regularization (C) — lower C penalizes large weights more, producing a smoother, wider-margin boundary; higher C fits the training points more tightly, sharpening the boundary but risking overfit to noise.
• Classify new note — drops a fresh, unlabelled document token into the space and shows which specialty region it lands closest to, with a confidence score derived from the margin.

Controls

• Vectorizer — switch between TF-IDF and raw term-frequency features.
• Classifier — switch between Logistic Regression and Naive Bayes.
• Regularization (C) — controls boundary sharpness / margin width.
• Show decision surface — toggle the coloured floor classification regions.
• 📄 Classify new note / ↺ Resample corpus — spawn a new document or regenerate the training scatter. Drag to rotate, scroll to zoom.

Did you know?

On real MTSamples-style corpora, TF-IDF + Logistic Regression typically reaches roughly 60-75% multi-class accuracy across a dozen or more specialties — respectable for a model with no understanding of medical semantics, but it still confuses clinically adjacent fields such as cardiology and cardiovascular/pulmonary notes that share heavy vocabulary overlap.