Speaker
Description
Alzheimer's disease (AD) is a progressive neurological disorder marked by memory deterioration and cognitive loss, impacting millions globally. Although AD is predominantly regarded as a brain disorder, increasing evidence indicates considerable peripheral modifications, especially within the microcirculatory system. In fact, research has shown that people with AD have platelet dysfunction and altered peripheral endothelial vascular responses. Infrared thermography (IRT) has emerged as a non-invasive method for evaluating peripheral microcirculation and autonomic functions by contactless detection of skin temperature. In fact, the temporal modulation of face skin temperature, especially at responsive areas like the nose tip, has shown sensitivity to fluctuations in autonomic states modifications. In this perspective, machine learning methods, especially deep learning (DL) with transfer learning (TL), allow detecting patterns in thermal data indicative of microcirculatory abnormalities.
This work aimed to develop a classifier using transfer learning with the EfficientNet-B0 architecture to differentiate AD patients from age-matched healthy controls (HC) using IRT spectrograms obtained from resting-state nose tip temperature measurements.
The study sample was composed of 12 AD patients (average age 71.6 ± 4.6 years; 7 men and 5 women) with moderate AD (according to DSM-5 criteria, MMSE > 20/30) and 15 HC (average age 69.3 ± 5.8 years; 10 men and 5 women). A FLIR SC660 camera (640×480 pixels, <30 mK sensitivity, 10 Hz sampling) was used to take thermal images of participants while they were resting for 5 minutes in a controlled setting (22-24°C, 45-50% humidity). The camera was 60 cm away from the individuals. After 15 minutes of acclimation, participants sat with their eyes closed while their face temperature was acquired. The nose tip was automatically tracked across frames, delivering a continuous temperature time series. Spectrograms of the nose tip temperature oscillations were computed and used as input of TL based on the EfficientNet-B0 model. To save learnt general features, the convolutional layers were frozen, while the last classification layer was modified to a binary classifier (AD vs. HC). Data augmentation based on random horizontal flipping, rotating by 15°, and color jittering was implemented. A 5-fold stratified cross-validation was used to partition the sample such that the classes were balanced. The Adam optimizer (learning rate: 10-4) was used for training, with cross-entropy loss for 10 epochs per fold (batch size: 8).
The EfficientNet-B0 model reached an overall accuracy of 78.7% across all folds, with a sensitivity of 86.7% and a specificity of 70.6%.
The findings indicate that time-frequency representations of spontaneous thermal oscillations recorded on the nose tip provide discriminative information for AD identification through DL methodologies. These results foster the employment of IRT combined with DL for the development of a screening method that is non-invasive, contactless, and comparatively cost-effective for clinical applications.
