An AI recognition method for children's clinical operative pain by skin potential (SP) signal

Children's clinical pain phenotypes are complex, and there is a lack of objective biological diagnostic markers and cognitive patterns. Detecting physiological signals through wearable devices simplifies disease diagnosis and holds the potential for remote medical applications. This research established a pain recognition model based on AI skin potential (SP) signal analysis. A total of 237 subjects participated in this study, comprising 152 boys and 85 girls, ranging in age from 2 to 16 years old. Initially, we preprocessed SP signals and built datasets for pain and non-pain conditions, including 195 pain and 97 non-pain samples. Then, we applied wavelet transform (WT) to capture the time-frequency characteristics of the signals and extract energy features and created a feature set comprising 30 features and selected 10 most relevant ones using the “SelectKBest” function. We compared six algorithms, optimized their parameters, and evaluated the stability and fitting performance of each algorithm. The random forest (RF) algorithm emerged as the best, demonstrating significant performance in pain recognition with an accuracy of 80.3% and a sensitivity of 92%. The SP signals generated by children of different genders, ages, and needling positions during indwelling needle puncture were accurately recognized. We developed a comprehensive SP recognition model, innovatively employing WT for SP signal analysis. This time-frequency analysis method, by preserving low-frequency features, is particularly suitable for SP signals. By combining pain monitoring with SP signals and ML, subjective pain experiences are transformed into quantifiable data, achieving high accuracy and real-time measurement capabilities. These advantages provide valuable technical support for clinical pediatric pain management. [Display omitted] •Real-Time Monitoring: Utilized wearable devices to monitor hospitalized children's skin potential (SP) signals in real-time.•Child-Specific Pain Recognition: Developed a child-specific pain recognition system, achieving 80.3% accuracy and 92% sensitivity, eliminating the need for interaction and enhancing objectivity and accuracy.•Wavelet Transform Analysis: Used CWT to capture SP signal time-frequency characteristics, revealing that over 99.5% of pain signal energy is within the 1.25Hz frequency band.•Optimized Feature Selection: Identified 10 key signal features, improving feature contribution rates and reducing computational complexity, leading to a