(G-258) Supervised Machine Learning based calprotectin monitoring by an Electrochemical Sweat Sensor

Introduction:: Wearable biosensors powered by machine learning have led to development of intelligent analytic methods that give insight into human health. Noninvasive continuous monitoring of biomarkers is especially important for disease state, severity and treatment management. Calprotectin is a state maker for inflammatory bowel disease and elevated levels are indicative of inflamed intestinal mucosal lining. Tracking calprotectin levels through periods of active, remission and relapse could provide valuable insight into gut health. Particularly in inflamatory bowel disease (IBD) where the current gold standard for measuring calrotectin levels is through a fecal sample. Fecal sample analysis is often accompanied by poor patient compliance and provide one discrete timepoint in a day as opposed to continuous monitoring. Sweat sensors provide a convenient method of monitoring as opposed to discrete measurement of the biomarker in standard fecal matter.  A passive electrochemical sweat sensor was developed along with a machine learning model that predicts calprotectin levels in real time, continuously using parameters of impedance and environmental factors.

Materials and Methods:: A supervised learning method, ensemble learning was implemented using the sensor parameters, time and environmental factors as the predictor variables and gold standard ELISA measurements of calprotectin as ground truth. Data was collected over a 48-hour time period for 13 healthy subjects. The sensor samples minute by minute values of calprotectin while ELISA measurements were obtained every 6 hours. The ground truth values were interpolated to yield minute by minute values to generate the training data. Bootstrapping was then used to reduce overfitting by diversifying the sample set. The trees were then aggregated together and parallel training was performed. The depth of the regression model was varied between 1 – 8 levels to optimize hyper parameters that yield the lowest residual squared error. The model was trained to predict concentration of calprotectin within the range of 0-6000 ng/ml every minute. 

Results, Conclusions, and Discussions:: Results and Discussion: Residual Mean Squared Error (RMSE) and R² was used to evaluate the results of the model. A goodness of fit of 0.92 and RMSE of 189 ng/ml was obtained. Figure 1(a) Showcases the workflow of the sensor as opposed to the current method used in the clinical framework. The training data is generated using ELISA and sensor parameters as input. The pre-processing pipeline comprises of interpolation, transformation and exploratory data analysis to remove highly correlate variables and handle missing values. Feature selection was used to build a robust model with the most significant features of magnitude of impedance, phase angle, relative humidity, skin temperature and time. A tree level of depth 8 was implemented as it

the lowest RMSE.  Figure 1(b) depicts the Bland Altman analysis, a validation metric that was conducted between the ML model and ground truth ELISA values with 95% confidence intervals. There is agreement between the two methods as most values are centered around the mean and lie between the acceptable bounds.

Conclusion: This study presents a data-driven implementation of a machine learning model in predicting calprotectin levels in human sweat. The model allows for a real time-based continuous tracking of the biomarker noninvasively within the trained physiological range. Tracking these levels could potentially give insight to flare-ups in inflammatory bowel disease.

Acknowledgements (Optional): : Department of Bioengineering, University of Texas at Dallas, EnLisense LLC

References (Optional): :

Jagannath, B., Lin, K.-C., Pali, M., Shahub, S., Abha Sardesai, Sriram Muthukumar, & Prasad, S. (2023). An observational study demonstrating the measurement, characterization and validation of expression of calprotectin in human sweat through a sweat wearable. 13, 100314–100314. https://doi.org/10.1016/j.biosx.2023.100314

‌

Jagannath, B., Lin, K.-C., Pali, M., Devang Sankhala, Sriram Muthukumar, & Prasad, S. (2020). A Sweat-based Wearable Enabling Technology for Real-time Monitoring of IL-1β and CRP as Potential Markers for Inflammatory Bowel Disease. 26(10), 1533–1542. https://doi.org/10.1093/ibd/izaa191

‌