Computational tools offer significant potential for improving diagnostics, particularly in point-of-care (POC) testing. Deep learning, for example, has shown promise in automating image analysis for disease detection. However, POC tests, often relying on low-cost materials and simple designs, frequently lack the accuracy of laboratory-based methods. Paper-based immunoassays, while affordable and user-friendly, often suffer from limitations in sensitivity, specificity, and reagent stability. The hook-effect, a phenomenon of competitive binding, can also lead to false results, particularly when the analyte concentration varies widely. This study addresses these challenges by integrating computational sensing with a paper-based VFA for hsCRP testing. hsCRP is a key biomarker for assessing cardiovascular disease (CVD) risk, with established clinical cut-offs for risk stratification. Accurate and precise hsCRP quantification is crucial, but existing paper-based systems often fall short. The researchers developed a deep learning framework to address these limitations, jointly optimizing the sensor design (multiplexed sensing membrane) and the quantification algorithm. Their previous work using neural networks for POC Lyme disease diagnostics provided a foundation, but this study uniquely focuses on precise protein biomarker quantification, incorporating fabrication information into the model, and extending the dynamic range through analysis of multiplexed immunoreaction spots.

The introduction extensively cites literature supporting the challenges of POC testing, particularly concerning paper-based immunoassays and the limitations of accuracy and sensitivity. It reviews the use of deep learning in medical image analysis and its potential applications in biosensing. Several papers are referenced regarding the hook effect and its impact on immunoassay results. Prior work by the research group on using neural networks for POC diagnostics is also highlighted, emphasizing the novel aspects of this current study: precise quantification (not just binary classification), inclusion of fabrication details in the model, and a wider dynamic range through multiplexing.

The study utilized a multiplexed paper-based vertical flow assay (VFA) for hsCRP testing. The VFA comprises stacked paper layers within a 3D-printed cassette, enabling uniform vertical flow of the serum sample across a nitrocellulose (NC) sensing membrane. The multiplexed sensing membrane incorporates up to 81 spatially-isolated immunoreaction spots, each defined by a unique 'spotting condition' (capture protein and buffer). A custom spot-assignment algorithm ensured even distribution of spotting conditions across the membrane to mitigate flow rate variations. An automated liquid dispenser deposited reagents onto the NC membrane, with multiple fabrication and reagent batches used to assess batch-to-batch variability. The assay procedure involved sequential injection of buffer, sample, and conjugate, followed by a 10-minute reaction time. A custom mobile phone reader captured images of the activated sensing membrane. Image processing software automatically detected and segmented immunoreaction spots, calculating normalized signals. The core of the methodology involved a deep learning-based framework. Machine learning was used in two stages: (1) spot selection to identify the most informative spots on the membrane, and (2) condition selection to determine the most robust combination of spotting conditions. A k-fold cross-validation approach was employed to optimize feature selection. The final CRP quantification algorithm was a fully-connected neural network trained using the entire training dataset. A separate classification algorithm identified samples outside the hsCRP range (acute inflammation). The inputs to the neural networks included normalized signals from selected spots and fabrication/reagent batch IDs.

The machine learning-based optimization selected 38 spots and 5 conditions. The cross-validation results demonstrated the effectiveness of feature selection, showing improved performance with fewer spots and conditions. Blind testing of 57 clinical samples showed excellent correlation with gold-standard hsCRP measurements (R² = 0.95). The average coefficient of variation (CV) was 11.2%, meeting or exceeding FDA criteria for hsCRP testing in various risk categories. The system achieved 100% accuracy in classifying samples as either within the hsCRP range or representing acute inflammation. Importantly, the multiplexed approach and computational analysis effectively mitigated the hook effect, allowing accurate measurement across a wide dynamic range, including samples with greatly elevated CRP concentrations. Incorporating fabrication information into the model significantly improved performance, reducing MSLE and CV while increasing R². The reagent cost was reduced by 62% using the optimized subset of spots and conditions.

The results demonstrate the power of a data-driven approach to sensor design and readout. The deep learning algorithms significantly improved quantification accuracy compared to standard regression methods. The inclusion of fabrication information enhanced model robustness and generalizability. The mitigation of the hook effect through multiplexing and computational analysis is a key achievement, expanding the dynamic range and avoiding false-negative results. The study highlights the potential of computational sensing to optimize POC tests, improving both performance and cost-effectiveness. This approach offers a path toward democratizing access to diagnostic testing, particularly in resource-limited settings.

This study presents a novel deep learning-enabled point-of-care sensing platform for hsCRP quantification, achieving high accuracy and precision in a clinical setting. The integrated framework, combining multiplexed VFA design with machine learning algorithms, effectively addresses challenges of traditional paper-based assays, including the hook effect. The cost-effective and portable nature of this technology holds significant promise for expanding access to CVD risk assessment. Future research could focus on exploring alternative neural network architectures, expanding the panel of biomarkers, and refining the cost function to better reflect clinical needs.

The study's reliance on a specific machine learning algorithm and architecture may limit generalizability. While the use of multiple fabrication and reagent batches helped assess variability, larger-scale manufacturing might reveal additional sources of error. The study's use of remnant human serum samples might not perfectly represent the full range of real-world samples. Further validation with a more diverse and larger sample size is needed to confirm the robustness of the findings in diverse populations.