Assessing pain in Alzheimer’s patients with communication limitations is a significant challenge, as traditional self-reporting methods are ineffective. In this paper, we introduce an AI-based, noninvasive system that detects pain levels by analyzing facial micro-expressions and physiological responses. Utilizing the Multimodal Intensity Pain (MIntPAIN) dataset, we extracted facial features using the InceptionResNetV1 model pre-trained on VGGFace2, producing 512-dimensional embeddings for each image, and then balanced the MIntPAIN dataset by equalizing the number of ‘no pain’ and ‘pain’ sequences. The proposed classification system includes a dimensionality reduction performed using Principal Component Analysis with a 0.99 threshold and a three-layered CNN-RNN model combining Conv1D and Bidirectional LSTM layers. The experimental evaluation shows the model achieved a training accuracy of 99.19% and a validation accuracy of 90.22%, demonstrating its effectiveness in accurately detecting pain levels. This research facilitates real-time analysis and operational efficiency for secure cloud deployment and offers a viable solution for pain assessment in nonverbal Alzheimer’s patients.