Computational Phenotyping of Unrecognized Prediabetes: A Reproducible Data-Analytics Development and Temporal Validation Protocol

Abstract

This protocol establishes a repeatable protocol for computational processing of public data to test the hypothesis that multiday wrist-accelerometry features can be used to distinguish prediabetes (based on multiple biomarker criteria) in adults not professionally diagnosed. Only deidentified NHANES 2011-2014 public use files on or before July 31, 2025, will be utilized in the study. These principal solutions are data analytic: deterministic wearable QA, predetermined circadian feature engineering, development-only preprocessing, regularized probability modeling, explicit leakage barriers, and frozen temporal validation. NHANES 2011-2012 is the development cycle and NHANES 2013-2014 is the untouched temporal-validation cycle. The Circadian Dysfunction Score (CDS) is the primary digital exposure, which is a combination of the lower relative amplitude, lower interdaily stability, higher intradaily variability and higher log transformed L5/rest-period activity. The main model is the ridge logistic regression model. Validation thresholds are not deployment thresholds, the primary incremental-validation estimand is the difference in specificity at 90% sensitivity between the nonlaboratory baseline model and the baseline-plus-CDS model at cycle 2013-2014. Separately, thresholds for cycle 90% and 95% sensitivity will be set at the development stage and then frozen and used without change across the rest of the cycle to measure the achieved sensitivity, specificity and referral burden. Secondary performance measures must include calibration intercept, calibration slope, Brier score, AUROC and precision-recall AUC. A separate survey-weighted epidemiologic layer will calculate prevalence and association for pooled 2011-2014 data based on the NHANES design variables and fasting-subsample weights. Prediabetes is defined by the thresholds set by the ADA: HbA1c and FPG, and prior recognition is defined by DIQ160. The wearable model is only a risk prioritization tool for biochemical screening and not a diagnostic substitute. There are no outcome estimates due to the fact that this manuscript is a protocol and a statistical analysis plan.