Deep learning-powered chemiluminescence vertical flow assay for high-sensitivity troponin detection at the point of care

Introduction: : Cardiac troponin I (cTnI) is one of the most specific and sensitive biomarkers for diagnosing myocardial infarction, a major contributor to global cardiovascular mortality. Centralized laboratory assays, while accurate and highly sensitive (down to a few pg/mL), are often inaccessible in low-resource settings due to their reliance on expensive equipment and trained personnel. To address this diagnostic accessibility gap, point-of-care testing (POCT) platforms have been proposed as cost-effective alternatives. However, most existing POCT devices fall short of achieving the ultra-low detection thresholds necessary for high-sensitivity cTnI (hs-cTnI) testing. In response, we developed a paper-based chemiluminescence vertical flow assay (CL-VFA) system powered by deep learning and computational imaging. The CL-VFA is designed to deliver laboratory-comparable performance in a low-cost, portable format, enabling rapid and accurate hs-cTnI quantification from 50 µL of serum samples in 25 min. The system integrates (i) a paper-based sensor using polymerized enzyme labels for high-intensity CL signal (Fig. 1a), (ii) a compact Raspberry Pi-based imager optimized for CL signal capture (Fig. 1b), and (iii) a fully automated deep learning pipeline for cTnI classification and quantification (Fig. 1c and d). By combining affordability, simplicity, and sensitivity, our platform provides a promising solution to the ongoing need for decentralized cardiac diagnostics—particularly in underserved and remote clinical environments.

Materials and

Methods: : The CL-VFA consists of two modular cartridges for sequential processing: the first for immunoassay and washing, and the second for CL signal generation and image capture. During the initial 20-min assay step, the serum sample is mixed with antibody-conjugated labels and loaded into a stacked paper structure in the sensor, where a sandwich immunoassay forms. After unbound materials are removed through buffer washing, the sensing membrane is transferred to the second cartridge. During the 5-min imaging step, the sensor is incubated with a CL reagent and imaged for 30 s using a Raspberry Pi-based portable imaging system. Four fully connected deep neural network (DNN) models then process the captured images (Fig. 1d). The first DNN model (DNNClassification) classifies the sample into one of three concentration categories ( < 40 pg/mL, 40–1000 pg/mL, >1000 pg/mL). Once the initial classification is complete, the sample is processed by one of the three cTnI quantification networks (DNNQ < 40, DNNQ40–1000, or DNNQ>1000). Training and validation were conducted using 144 serum samples from 72 patients, while performance evaluation was done with 66 samples from 34 patients, none of which were used in the training or optimization phases.

Results, Conclusions, and Discussions:: Our CL-VFA platform demonstrated a detection limit of 0.16 pg/mL and a dynamic range spanning six orders of magnitude, with an average coefficient of variation below 15%. These metrics meet the established criteria for high-sensitivity cTnI assays and support its clinical relevance in early myocardial infarction detection. In blinded testing with 66 serum samples, the deep learning-based classification model achieved a diagnostic accuracy of 95.5%, correctly sorting samples into appropriate cTnI concentration categories (Fig. 1e). Concentration inference from the quantification networks showed strong correlation (Pearson’s r = 0.984) with reference values obtained using an FDA-cleared laboratory analyzer (Fig. 1f), validating the analytical robustness of the platform. Beyond performance, the system’s affordability and simplicity provide distinct advantages for practical applications. Our CL-VFA costs $4.25 per test, and the mobile reader is $222 per prototype, presenting a cost-effective alternative to laboratory testing. This enables deployment in low-resource or decentralized clinical settings, expanding access to high-quality cardiac diagnostics in environments where it is most critically needed.

In conclusion, the deep learning-enhanced CL-VFA platform represents an important step toward democratizing access to high-sensitivity cardiac biomarker testing. By integrating paper-based assays, computational imaging, and neural network-driven analysis, this work addresses the technical and logistical challenges that POCT faces in cardiovascular care. Future development will focus on incorporating multiplexed cardiac biomarker detection and expanding clinical validation using whole blood-based samples.

Acknowledgements and/or References (Optional)::