PathoSafe: Sub-1 KB Out-of-Distribution Rejection and Cluster-Aware Certification for Point-of-Care Pathogen Biosensing
A point-of-care pathogen-microscopy readout must signal when a sample falls outside what its detector was trained on. PathoSafe is an 816-byte int8-projection low-rank Mahalanobis out-of-distribution (OOD) head reading a frozen detector’s penultimate feature. It reaches a cross-site area under the receiver operating characteristic curve (AUROC) of 0.980 and multi-source AUROC of 0.964 (parasitic-egg-dominated), within 0.01 of a 16,896-byte full-precision dense baseline, and adds a measured 30.76 µs and 816 bytes of read-only memory (ROM) constants on an STM32H743 development board, so abstention adds only a small measured cost on-chip. The advantage is the trunk feature rather than the integer form. Certification is the harder problem and carries our central result: slide clustering silently breaks the standard independent-sample certificate, since distribution-free risk control assumes exchangeable samples that patch-clustered medical data do not supply. On a degraded-input risk certificate, the naive cell-level version holds on only 20% of leakage-free re-splits, whereas the correct slide-level one is valid at 0.266 with its width set by the slide count rather than by any bound we evaluate. The deployed threshold inherits this less severely. A Hoeffding–Bentkus budget gives an illustrative count of 38 independent slides under the stated assumptions (35 when five seeds are used), which is about 1.3× what this benchmark provides. We release that leakage-free slide-disjoint benchmark: an in-distribution malaria task with cross-site (BBBC041) and multi-source (SIPaKMeD, parasitic egg, white blood cells) regimes.
Authors
- Longquan Chen (ORCID: https://orcid.org/0000-0002-6785-5914)
- Gaoming He
- Qihuang Gao
- Mengdi Hou
- Bitie Lan
- Jianbo Huang (ORCID: https://orcid.org/0009-0009-3549-9036)
Institutions
- Guilin University of Aerospace Technology (CN)
- Wuzhou University (CN)
- Guilin University of Electronic Technology (CN)
Publication Details
- Journal
- Biosensors
- Published
- 2026-09-17
- DOI
- https://doi.org/10.3390/bios16090518
- Primary Topic
- Cell Image Analysis Techniques
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- Science and Technology Department of Guangxi Zhuang Autonomous
- Wuzhou University