TimeSight: Kinetics-aware detection for digital nucleic acid amplification via spatiotemporal convolutional learning

Digital nucleic acid amplification enables absolute molecular quantification by partitioning samples into thousands of micro-reactions, yet current analysis pipelines classify each partition from a single endpoint fluorescence image, discarding the temporal dynamics that distinguish true amplification from artifacts. This limitation is acute in portable point-of-care testing (POCT) instruments, whose simplified optics and low-cost sensors produce images with non-uniform illumination, elevated background noise, and low signal-to-noise ratios. This study presents TimeSight, an end-to-end spatiotemporal detection framework for time-resolved fluorescence image stacks acquired during digital recombinase polymerase amplification (dRPA) on a portable POCT instrument. Built upon factorised three-dimensional convolutions with a learnt temporal collapse mechanism, TimeSight jointly encodes spatial morphology and amplification kinetics without partition segmentation, spatial registration, or manual threshold tuning. On a held-out test set of 8248 ground-truth partitions, TimeSight achieved the highest F1 (0.968) and precision (0.977) among all single-frame and multi-frame baselines while adding <2% parameters. Feature activation analysis confirms that the network suppresses temporally invariant artifacts indistinguishable from true positives in endpoint images, maintaining consistent performance across normal, low-SNR, and artifact-dominated conditions. TimeSight matched reference concentrations across four orders of magnitude and agreed with commercial digital PCR on genuine S. aureus DNA, with edge deployment confirming portable feasibility at about 113 ms per stack. These results indicate that temporal amplification kinetics can compensate for the spatial information lost to hardware miniaturisation, lowering the barrier to point-of-care absolute quantification. TimeSight offers a spatiotemporal detection paradigm for digital nucleic acid quantification in resource limited settings.