Title : Dementia AI: A low-cost, privacy-preserving mobile application for automated dementia screening using multinomial logistic regression and computer vision
Abstract:
Dementia is a global health challenge, with over 55 million diagnosed cases and a significant portion of undiagnosed individuals, particularly in low-income countries. Early detection is critical for effective intervention, yet traditional screening methods like the Clock Drawing Test (CDT) are prone to bias due to manual scoring. This study introduces *Dementia AI*, an AI-based mobile application designed to automate dementia screening using multinomial logistic regression, computer vision, and machine learning. The application leverages a dataset of 16,209 CDT images, trained using Create ML and Scikit-learn, to classify clock drawings into ordinal scores (0-5) indicating dementia likelihood. The iOS-based app integrates CoreML for on-device processing, ensuring privacy and eliminating the need for internet connectivity.
Dementia is a global health challenge, with over 55 million diagnosed cases and a significant portion of undiagnosed individuals, particularly in low-income countries. Early detection is critical for effective intervention, yet traditional screening methods like the Clock Drawing Test (CDT) are prone to bias due to manual scoring. This study introduces Dementia AI, an AI-based mobile application designed to automate dementia screening using multinomial logistic regression, computer vision, and machine learning. The application leverages a dataset of 16,209 CDT images, trained using Create ML and Scikit-learn, to classify clock drawings into ordinal scores (0-5) indicating dementia likelihood. The iOS-based app integrates CoreML for on-device processing, ensuring privacy and eliminating the need for internet connectivity.
Dementia AI represents a significant improvement over prior research in several key areas. Unlike existing solutions that rely on pre-trained Convolutional Neural Networks (CNNs) or digital pen technology, which are limited to binary outputs ("pass" or "fail") or require specialized hardware, Dementia AI provides an ordinal scoring system (0-5) that quantifies the severity of cognitive impairment. This nuanced approach allows for more precise screening and early detection of mild cognitive decline, which is often missed by binary systems. Additionally, while prior methods often depend on cloud-based processing or expensive MRI imaging, Dementia AI operates entirely on a smartphone, using only a pencil, paper, and the device's camera, making it low-cost, accessible, and privacy-preserving. The app’s ability to process images locally without internet connectivity further enhances its usability in resource-limited settings.
The developed model achieved a 95% accuracy in classifying CDT images, with the Create ML model outperforming Scikit-learn in F1 score (0.95 vs. 0.86). The app’s robustness was tested under varying lighting conditions, revealing no significant difference in accuracy between low and medium brightness settings (p = 0.348). This demonstrates the model’s reliability across environments. Dementia AI offers a low-cost, privacy-preserving, and user-friendly solution for dementia screening, with potential for remote use and early intervention. Future work includes expanding to Android platforms, exploring additional algorithms, and integrating multimodal AI approaches for enhanced screening capabilities. This project highlights the potential of mobile AI applications in healthcare, particularly for underserved populations.