Alzheimer’s disease (AD) is a neurodegenerative disease that causes a decrease in cognitive and motor skills. Early diagnosis of the disease helps reduce the financial burden and facilitate care. Researchers are increasingly exploring machine learning techniques for the early detection of AD. Existing diagnostic methods which are often costly and invasive, create a need for accessible, non-invasive screening approaches. This study investigates the feasibility of using motion data collected from a handheld smart ball device, PALLO, to distinguish between healthy individuals and those diagnosed with AD. Ten participants (6 healthy, 4 with AD) performed three types of movement tasks: walking, hand-raising, and memory-based movement in a controlled environment. Two machine learning approaches were implemented: a hybrid Convolutional Neural Network Long Short-Term Memory (hybrid CNN- LSTM) and a feature-based machine learning model (random forest classifier, gradient boosting, logistic regression, and support vector machine). The CNN–LSTM achieved the best classification performance, with 96% accuracy and an F1 score of 0.97. Among feature-based models, gradient boosting achieved the highest accuracy of 81%. Both approaches were more effective at classifying AD participants than healthy controls. Despite the small dataset, these findings demonstrate the potential of motion-sensor data combined with machine learning techniques for early AD screening. The proposed approach is simple, non-invasive, and could be extended to real-world settings or adapted for other conditions affecting motor function.