"An Unsupervised Approach to Distinguishing Dyslexic from Non-Dyslexic Writing Achieving up to 94% Purity: A Comprehensive Clustering Analysis with Feature Subset Evaluation "

Dyslexia is a learning difference affecting the way individuals recognize, spell, and process written language. While supervised machine learning techniques have shown promise in identifying dyslexia markers, unsupervised methods offer an alternative viewpoint by allowing hidden structures to emerge from data without requiring explicit labels during clustering. This paper presents a comprehensive, multi-stage unsupervised analysis aimed at discerning characteristics of dyslexic and non-dyslexic writing samples. We explore a diverse set of 13 handwriting and text-based features—including metrics for letter formation errors, baseline disruptions, spelling scores, and phonetic correctness—and systematically apply and compare four clustering algorithms: K-Means, Agglomerative Clustering, DBSCAN, and HDBSCAN. We then evaluate each method internally using Silhouette Score and externally using Purity, Adjusted Rand Index (ARI), and V-Measure against known labels (dyslexic vs. non-dyslexic). Our data encompass 100 writing samples, evenly split between dyslexic and non-dyslexic participants. Results indicate that K-Means with k=2 clusters often achieves strong internal cohesion (Silhouette ≈ 0.445) and good external alignment (e.g., ARI up to 0.543–0.772, depending on the feature subset). Agglomerative Clustering sometimes surpasses K-Means in ARI and Purity, although it can have a slightly lower Silhouette. DBSCAN and HDBSCAN are effective at discovering multiple clusters, with some runs achieving Purity of 0.90–0.91, yet their Silhouette Scores are generally lower. Crucially, a feature subset comprised of (lfe_value, poc_value, ds_spelling_score, ds_percentage_corrections) raises ARI to 0.772 and Purity to 0.94, demonstrating the value of well-chosen variables. We include detailed tables of clustering performance and six figures illustrating the elbow method, silhouette analysis, parallel coordinates, pairwise distributions, and PCA visualizations. Our findings highlight the significance of unsupervised exploration in dyslexia-related research, shedding light on how various features—and subsets of features—can elucidate distinct patterns in writing behaviors.