Alzheimer's disease (AD), an incurable disease, poses a major health problem. It is important to identify patients with mild cognitive impairment (MCI) and early AD. Clock rendering test (CDT) is an effect way to screen AD patients quickly in the community. However, the current CDT methods require specific equipment to collect features, and the existing prediction models are inefficient in early warning of MCI. To solve the problem, this paper replaces digital pen with fingertip interaction, and proposes an early warning model for AD early dCDT images based on ResNet50. The dCDT tests were carried out on normal cognitive elderly, MCI patients and mild AD patients, and the results were used to verify the analysis and classification ability of the ResNet50-based early AD prediction model, in contrast to the clock score-based early AD prediction model. The comparison shows that the ResNet50-based early AD prediction model is efficient in early warning than the other model, and is suitable for large-scale screening of AD patients in the community, in the absence of doctors. [ABSTRACT FROM AUTHOR]