Uncovering subgroup heterogeneity in dyslexia intervention outcomes using explainable machine learning.

Abstract

Identifying heterogeneity within literacy intervention outcomes can inform more targeted strategies for dyslexia remediation. Based on prior work that used machine learning to predict literacy intervention responders and non-responders at baseline, this study uncovered 2 subgroups of responders and 4 subgroups of non-responders based on feature importance profiles of 132 students who were diagnosed with dyslexia. Their average age was 8.2 years at baseline (98.08 months). They received literacy intervention for an average of 33.62 months. Intervention subgroup membership predicted long-term literacy development of a subsample of 41 students 68.2 months post intervention. The baseline scores, top predictors and long-term literacy outcomes were profiled for each subgroup. 'High Baseline Responders' and 'Low Baseline Responders' characterized the two responder subgroups. 'Total Regression Non-Responders', 'Self-Taught Non-Responders', 'Phonics-Receptive Non-Responders' and 'Lowest Baseline Non-Responders' characterized the four non-responder subgroups. Overall, our research found that clustering could identify subgroup profiles and demonstrates the predictive potential of longitudinal literacy development.