Highly accurate heart failure classification using carbon nanotube thin film biosensors and machine learning assisted data analysis

Abstract

While carbon nanotube (CNT)-based biosensors have been demonstrated to achieve high sensitivity and fast response, their intrinsic variation has been a challenge for the robustness of CNT-based diagnostic systems and greatly hindered their real-world clinical applications. The solution through conventional analytical methods demands advanced electronics and computing systems, which limits the feasibility of CNT-based biosensors in near-patient settings considering the device size, cost, and usability. Here, we present a machine learning (ML) assisted solution by analyzing electrochemical impedance spectroscopy (EIS) data generated by CNT thin film (CNT-TF) biosensors in response to B-type natriuretic peptide (BNP) for the diagnosis and management of heart failure (HF). We trained a deep neural network to predict the BNP level, where a K-fold cross-validation was used to optimize the network structure. This algorithm circumvents the error-prone analytical methods that require calibration curves, and significantly enhanced the reliability of diagnostic results without the need for advanced hardware or major design change. We first demonstrated the feasibility of our method using EIS data acquired in phosphate-buffered saline (PBS) and then applied the algorithm to venous blood samples collected from HF patients. Overall, an out-of-sample prediction accuracy of 80.0% (and an in-sample accuracy of 97.2%) was achieved for the PBS data and 80.0% (95.6%) for the clinical data, which shows that ML-assisted algorithms greatly improve the performance and applicability of CNT-TF biosensors in point-of-care testing. In general, this work demonstrates the great potential of ML in navigating challenges in sensor data analysis, especially in situations involving noisy data and complex data structures.