A Data-Driven Framework for Precision Hydroxyurea Therapy in Pediatric Sickle Cell Anemia Using Microsensor Arrays and Machine Learning
Year: 2026
Type: preprint
Abstract: This manuscript presents a structured data-driven framework for the precise quantification of hydroxyurea in pediatric sickle cell anemia patient samples. By integrating multidimensional patient clinical metadata with temporal alignment of dosing and sampling, along with highly overlapping and masking electrochemical signatures originating from blood samples, we developed a robust approach to mitigate signal noise and artifacts in sensor data, enabling reliable quantification in a small clinical cohort (n = 50). To enhance predictive accuracy, multiple feature-engineering strategies were employed, incorporating diverse representations of electrochemical variability, including averaged repeat signals across samples and electrodes, frequency-domain features, and normalized signal integrals. Benchmarking of multiple machine-learning models using 50 iterations of nested cross-validation, together with preprocessing via nonlinear kernel cosine principal component analysis, showed that Elastic Net performed best for Repeat-2 across samples and electrodes (R2) and R2-50% configurations, resulting mean RMSE values of 69.56 µM (5.29 µg/mL) and 66.53 µM (5.06 µg/mL), respectively. To address the loss of temporal electrochemical variability during preprocessing (from the full 21-electrode dataset to a refined 13-electrode), we further implemented a reservoir computing approach combined with Elastic Net regression, which reduced the RMSE to 62.37 µM (4.74 µg/mL) without evidence of systematic overfitting. A comprehensive optimism-bias assessment indicated that signal-averaging approaches were more stable (mean bias = 1.64 µM for the R2 dataset) than frequency-domain features. In parallel, the classification pipeline for patient stratification achieved an accuracy of 0.82 and an F1-score of 0.79. Building on the reference study by Shukla et al. (Anal. Chem. 2025, 97, 44, 24399-24414), this work emphasizes model robustness, advanced electrochemical feature engineering, and introduces patient demographic information as an additional factor to improve generalization. Overall, these findings establish a validated framework for AI-enabled point-of-care systems for hydroxyurea therapeutic drug monitoring in pediatric sickle cell anemia and provide a foundation for more precise, data-driven clinical decision-making across diverse healthcare settings.
Source: ChemRxiv
Institutions: Ben-Gurion University of the Negev, University of Cincinnati
Language: English
Subfield: Electrical and Electronic Engineering
Field: Engineering
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7162316988