Impact of AI-Augmented Histopathology Review on Next-Generation Sequencing (NGS) Success

Background
Insufficient nucleic acid quantity (QNS) can compromise NGS success (QNS rates can be > 20%), necessitating re-biopsies and delaying therapy. While interventions like extended de-crosslinking (EXT) can rescue samples with low total nucleic acid (TNA) yield tissues, they prolong processing times, demanding a precise method to identify samples requiring such processing. We developed an AI system, Paige Predict (PP), that analyzes digitized H&E slides to predict NGS QNS and recommend tissue input quantity. This study evaluates PP in two contexts: (1) optimizing internal lab workflows by selectively routing high-risk samples to EXT, and (2) exploring its utility in an external setting, Cedars-Sinai (CS), to triage samples for comprehensive Tempus genomic profiling (CGP), requiring > 50ng input, vs targeted low-input assays ( > 5ng) performed at CS.

Methods
We conducted a validation study in the Tempus lab comparing a baseline period prior to introduction of EXT (July 2024 – Sept 2024 n = 17,026) against an intervention period (Oct 2025 – Nov 2025, n = 12,975) where PP automatically routed at-risk samples to EXT. To evaluate external utility, we also performed a retrospective analysis on a cohort of CS pts. We assessed PP’s ability to predict CGP QNS, and compared it against pathologist assessment for inter-institutional sample referral.

Results
In the validation study, PP-guided EXT routing reduced joint DNA+RNA QNS rates by 19.6%, with a number needed to test (NNT; 1/Absolute Rate Reduction) indicating that for every 40.2 pts, one received a result that would otherwise not have. RNA QNS rates reduced by 15.9% (NNT 63.7) relative to the baseline period. This reduction was achieved while decreasing net tissue input by 16.7% and increasing TNA yields in the optimal range (100–1500ng) by 19.0% (NNT 10.1). Similar benefits were observed in the NSCLC subset. In the external CS cohort, PP could have triaged 74% of samples that failed Tempus CGP to low input sequencing at CS. Conversely, PP indicated that over 70% of samples sequenced at CS could have succeeded with Tempus CGP.

Conclusions
PP significantly reduces NGS failure rates and improves tissue stewardship by deploying rescue workflows only when necessary. These analyses lay the groundwork for a future multi-institution prospective trial and establish PP as a scalable AI tool to amplify access to precision oncology.