Spiking neural network-inspired streaming digital droplet PCR detection algorithm
TL;DR
A neuron-dynamics-inspired algorithm for droplet event detection and feature extraction that emulates membrane potential integration through pulse accumulation, introduces a refractory period to suppress consecutive false positives, and combines multi-condition peak verification to achieve robust event localization is proposed.
Abstract
In flow cytometric detection of digital droplet PCR, pulse signals from photomultiplier tubes are susceptible to noise, baseline drift, and overlapping adjacent peaks, posing challenges for reliable droplet event recognition. Inspired by the dynamics of neurons in spiking neural networks, this study proposes a neuron-dynamics-inspired algorithm for droplet event detection and feature extraction. The algorithm emulates membrane potential integration through pulse accumulation, introduces a refractory period to suppress consecutive false positives, and combines multi-condition peak verification to achieve robust event localization; simultaneously, it extracts multidimensional feature vectors including pulse accumulation values derived from the spiking neural network internal state. Evaluated on experimental PMT signals across multiple flow rates and droplet sizes, the algorithm reliably resolves dense events with overlapping waveforms. Under adaptive threshold selection, the event-level F1-score reaches above 99.9% across all tested throughput conditions, with a distinct advantage in reducing false positives compared to sliding-window local thresholding at high throughput. This work provides a new bio-inspired signal processing framework for digital droplet PCR analysis with demonstrated engineering practicality.