Machine Learning to Improve the Access to the Clinical Pathway for Children with Specific Learning Disabilities

A prompt diagnosis of specific learning disabilities (SLDs) is prevented by an overwhelmed healthcare system. As teachers lack clinical preparation, school-based screening needs to be improved. This work proposes methods to (1) identify children's profiles, (2) select children who need a visit, and (3) provide a better understanding of the characteristics connected with the start of the clinical pathway. We analyzed data from 364 children referred to clinical consultation. Starting from a 96-item screening questionnaire filled in by teachers at school, we computed a severity score for 19 different sub-domains of learning leveraging item response theory. Then, we performed cluster analysis with K-means to segment the population according to children's capabilities. For each cluster, we leveraged leave-one-out on balanced outcomes (clinical pathway VS school training) with different machine learning models, and we leveraged Shapley values to explain the results. Cluster analysis revealed two children's profiles, grouped by severity. Though, the proportion of children who started the clinical pathway was not statistically different. Indeed, also children with less difficulties should be taken into consideration, as they may suffer from SLDs without comorbidities. As for the classification, median area under the precision-recall curve was 0.96 for one cluster with a Support Vector Classifier (SVC), and 0.69 for the other cluster with Naive Bayes (NB). Between-cluster differences in performance suggest different degrees of complexity in children's profiles. Yet, also the latter can be considered a good results, considering the heterogeneity of data creators. Shapley values revealed that the SVC on the first cluster tended to rank children by severity, whilst NB on the second cluster shows that the difficulties can interact in a more complex way. This work represents a step forward in the management of SLDs, from an early and preclinical setting.Clinical relevanceThis work provides methods to get insights on the reasons for referring children with specific learning disabilities to the clinic.