Waveform-aware deep learning for quantitative agglutination on a low-cost lab-in-tubing microfluidic platform

To enable quantitative agglutination readout using an ultra-low-cost (parts cost < $40) lab-in-tubing optical platform. Droplet waveforms were recorded with dual infrared line-break sensors from biotinylated BSA mixed with streptavidin-coated beads at nine levels (0, 0.0128–1000 μg mL-1; N = 223 droplets, N = 211 after QC). We apply waveform-aware preprocessing (i.e., preprocessing operations defined on the full time-series waveform that preserve concentration-dependent morphology rather than reducing the signal to a single scalar) (length normalization, Hampel spike suppression, offset-drift compensation, quality-control filtering, and blank (0 μg mL-1) reference subtraction) and train a one-dimensional convolutional neural network (1D-CNN) to regress log10(concentration). Results: On a stratified 80/20 validation split, the model achieved MAE = 30.49 μg mL-1 with sub-μg absolute errors at trace-to-mid concentrations and larger but scale-appropriate deviations near the apex (~8 μg mL-1) and in the prozone. Leave-one-concentration-out (LOCO) experiments show that errors increase primarily at omitted calibration anchors, motivating a minimal three-anchor scheme (low, apex, high/prozone) for stable interpolation. Conclusion: Coupling inexpensive optics with waveform-level modeling supports quantitative agglutination estimation across nearly five orders of magnitude (0.0128–1000 μg mL-1). Multi-day, multi-device, and matrix-effect validation remain future work. LOCO is a special case of leave-group-out cross-validation [26], and we use it here to emulate missing calibration anchors [27]. • Waveform-aware preprocessing isolates agglutination morphology from low-cost optical signals. • 1D-CNN regression estimates concentration across 0.0128–1000 μg mL-1 (nearly 5 orders). • LOCO analysis motivates a three-anchor calibration strategy (low, apex, high/prozone).