Askisi-MD: Development and Validation of a Web-Based Neuropsychological Screener for Mathematical Learning Difficulties Using Explainable Machine Learning.

Abstract

Background: Mathematical learning difficulties (MLD) are associated with weaknesses in both domain-specific numerical abilities and domain-general neurocognitive processes. In this study, we developed and validated a web-based neuropsychological screener, Askisi-Mathematical Difficulties (MD), that integrates cognitive control and mathematics-related tasks for the first-line identification of children at risk for MLD was developed and validated.

Methods: We enrolled 564 children, including 282 children with an established MLD diagnosis recruited from a state diagnostic center and 282 typically developing children pair-matched for age and gender. Askisi-MD consisted of six tasks, including three cognitive tasks assessing inhibitory control, visual pattern recognition, and visual working memory, as well as three mathematics-related tasks assessing mental calculation verification, mathematical terminology comprehension, and arithmetic word problem operation selection. Accuracy scores and latency indicators were analyzed. Differences among groups were examined using Welch's independent-samples t-tests, Benjamini-Hochberg False Discovery Rate (FDR) correction, and Hedges' g effect sizes. Internal coherence was examined using McDonald's Ω. Discriminant validity was evaluated via receiver operating characteristic (ROC) analysis and through the use of classification indices. Supervised machine learning models, including Support Vector Classifier, Random Forest, and eXtreme Gradient Boosting (XGBoost) classifiers, were trained and cross-validated, while model interpretability was examined via SHapley Additive exPlanations (SHAP) analysis.

Results: Children with MLD showed significantly poorer performance across all six task-score indicators after FDR correction, with moderate-to-large effect sizes. Minute-based task-completion time differences were more selective. Children with MLD showed significantly longer completion times on the Mental Calculation and Arithmetic Word Problem tasks, whereas completion-time differences were not significant for Visual Pattern Recognition Abilities or Mathematical Terminology Comprehension. McDonald's Ω indicated modest internal coherence for the neuropsychological task-score composite and the total task-score composite. The strongest internal coherence was observed for the mathematical processing-efficiency composite. ROC analysis supported strong case-control discrimination, and classification indices indicated balanced group classification. Across the cross-validation folds, XGBoost achieved the highest mean accuracy, precision, and F1 score, while XGBoost and Random Forest demonstrated the same mean recall. XGBoost also showed strong discriminative performance, as indicated by the cross-validated area under the receiver operating characteristic curve (ROC-AUC). SHAP analysis indicated that Arithmetic Word-Problem Task Latency, Arithmetic Word Problem performance, Mathematical Terminology Comprehension, Mental Calculation, and Go/No-Go error score had greater influence on XGBoost model predictions than the remaining task indicators.

Conclusions: Askisi-MD provides a neuropsychologically informed web-based approach to MLD screening by jointly sampling mathematical performance and supporting cognitive control functions. The composite-level internal coherence findings support the use of Askisi-MD as a multidomain screener rather than as a set of independent mathematics subscales. Its explainable machine learning framework further enhances model interpretability by identifying the task-level indicators most strongly contributing to classification outcomes. Askisi-MD thus offers promise as a first-line screening tool to support referral decisions, rather than a stand-alone diagnostic instrument.