Prolonging Sample Stability with Machine Learning Expands the Catalog of Tests Available for Mail-In Capillary Blood Collected At-Home

Self-collection of biospecimens at-home, without specialized equipment or training, are increasingly being adopted in clinical practice due to convenience and patient preferences. However, sample instability during shipment means that remote access to common blood tests remains challenging. We hypothesized that the inaccuracy and imprecision in test results that develop because of sample instability could be modeled and controlled using knowledge of transit conditions captured by environmental sensors. We subjected 2685 blood samples from 65 participants to temperature cycles derived from real-world transit conditions. Training a model called Remote Control to predict change enabled accurate calibration of test results to approximate the time zero value at the point of collection, despite sample degradation. With calibration, unprocessed whole blood could be transported, for up to 9 days under ambient conditions and exposed to temperatures between 3.4 and 47.4 °C. Under these conditions, agreement with CLIA TEa ranged between 98.1 and 100%, with a |%bias| of 0.1-1.6%, a %CV of 2.2-4.9%, and a minimum sigma metric between 3 and 8.8σ for lipids (Cholesterol, HDL, LDL, Triglycerides, APO-A1, and APO-B). Performance was linear across measurement intervals (R², 0.929-0.995), accurate at medical decision levels (MCC, 0.808-0.981), with uniform model error under differing laboratories, equipment systems, reagent chemistries, patient demographics, health status, and medication status. Importantly, performance generalized to an independent cohort of participants collecting capillary blood. This software-based approach to prolong sample stability expands the catalog of tests available at-home, supports decentralized clinical trials and public health screening initiatives, and improves how healthcare can be affordably delivered at-home and on the patient's terms.