Automated Infrared Thermography and Machine Learning for Detecting Physiological Thermal Patterns in Lumbar Radiculopathy: An Exploratory Pilot Study

Background/Objectives: Biomimetic and bioinspired approaches seek to translate biological processes and physiological responses into measurable signals that can inform technological and computational systems. In lumbar radiculopathy, nerve-root dysfunction may affect peripheral sympathetic regulation and alter the physiological mechanisms governing cutaneous temperature distribution. Infrared thermography provides a non-invasive means of sensing these naturally occurring thermal responses, while machine learning enables computational recognition of complex patterns within the resulting physiological signals. From this bioinspired perspective, this exploratory pilot study evaluated the discriminatory performance of infrared thermography combined with machine-learning models for lumbar radiculopathy, using electroneuromyography (ENMG) as the reference assessment. Methods: Forty-four individuals attending a neurosurgery outpatient clinic were included: 14 asymptomatic controls (mean age: 38 ± 14 years) and 30 symptomatic participants (mean age: 61 ± 12 years), of whom 25 had abnormal and five had normal ENMG findings. Infrared images were acquired using a FLIR T540 camera, and the lower limbs were segmented into 39 regions of interest. Temperature differences were analyzed using three predictive models: random forest, neural network, and logistic regression. Results: Logistic regression was used for the detailed root-specific classification analysis. The areas under the receiver operating characteristic curve (AUCs) were 0.47 for L4, 0.50 for L5, and 0.75 for S1. Sensitivity, specificity, and accuracy were 66.7%, 40.0%, and 50.0% for L4; 50.0%, 0.0%, and 25.0% for L5; and 50.0%, 100.0%, and 87.5% for S1, respectively. Among the evaluated imaging perspectives, the posterior lower-limb view provided the greatest discriminatory information. Conclusions: Automated analysis of infrared thermal patterns showed limited discriminatory performance for L4 and L5 radiculopathy but higher preliminary discriminatory performance for S1. From a biomimetic perspective, this approach explores how naturally occurring physiological thermoregulatory responses associated with neural dysfunction can be sensed and computationally translated into recognizable patterns. The findings provide preliminary feasibility evidence for integrating non-invasive physiological sensing with machine learning as a bioinspired strategy for characterizing lumbar radiculopathy-associated thermal patterns. Larger, prospectively designed studies are required to validate these findings, address potential confounding factors, and determine the clinical applicability of this approach.