Advancing ADHD diagnosis: using machine learning for unveiling ADHD patterns through dimensionality reduction on IoMT actigraphy signals

Mental health is an integral component of overall well-being, profoundly influencing the lives of individuals, families, and communities worldwide. As our understanding of mental health has evolved, so too has our awareness of the myriad conditions that impact it. Attention-Deficit Hyperactivity Disorder (ADHD) emerges as a substantial and intricate mental health concern, classified as a neurodevelopmental disorder with a pronounced impact on an individual’s capacity to concentrate, manage impulses, and modulate their activity levels. That is why the actigraphy measurements from wrist actigraphy provide a non-invasive monitoring method and diagnosis too. This paper proposes a methodology employing dimensionality reduction for feature extraction from time series actigraphy data signals and Machine learning (ML) classifiers for classifying control subjects from ADHD patients. Three dimensionality reduction techniques UMAP (Uniform Manifold Approximation and Projection), t-SNE (t-Distributed Stochastic Neighbour Embedding), and PCA (Principal Component Analysis) were applied to the actigraphy sample of length 1440 (24-h recordings). A Set of 10 ML classifiers (Medium Tree, Efficient Logistic Regression, Gaussian Naïve Bayes, Quadratic SVM, Cubic SVM, Fine kNN, Ensemble Boosting, Ensemble bagging, Medium NN, and Kernel SVM) were trained on three reduced feature sets. On UMAP features Cubic SVM outshines with remarkable accuracy of 99.2%, on t-SNE features Medium Tree classifier achieves an impressive accuracy of 98.3%, and on PCA. The Medium Tree classifier stands out with exceptional accuracy at 99.7%. Quadratic SVM shows Consistent Performance by maintaining the highest accuracy across UMAP (97.5), t-SNE (89.9), and PCA (98.8). This research contributes to advancing our understanding of ADHD through objective, data-driven methodologies.