Year: 2026
Type: preprint
Abstract: PCR underpins hundreds of millions of clinical tests each year. It outputs a complex curve that approximates the concentration of a target genetic sequence in a sample. To classify samples as positive, labs then apply simple threshold rules to the curve, discarding most of its rich signal. This approach could introduce misclassifications, but because PCR is considered the gold standard for clinical samples, there is no ground truth against which to measure accuracy. In a set of 24,756 quality-control samples from a federally-certified lab for which ground truth is known, we show that threshold-based approaches to classifications are suboptimal. The lab’s threshold achieved zero false positives but a 4.35\% false-negative rate. This was not simply a poor choice of threshold: simulations show no threshold can cut false negatives below 1\% without dramatically raising false positives. We train a deep learning model (SPARK) that outperforms threshold rules, and specifically achieves very low false negatives. It performs particularly well on highly diluted samples, the hardest for threshold rules to flag. Finally, in inconclusive clinical samples, SPARK accurately forecasts the results of retesting, using only data from the initial sample. This could reduce retesting by up to 56\% while maintaining accuracy.
Source: Research Square
Language: English
Subfield: Infectious Diseases
Field: Medicine
Domain: Health Sciences
Open Access status: green
OpenAlex ID: W7130640688