ウェアラブル活動と多因子遺伝的リスクを用いた2型糖尿病の発生率予測：All of Usにおける生存モデリング研究

Evidence is limited on whether daily step counts are uniform across individuals or whether combining them with genetic risk improves prediction of type 2 diabetes (T2D), both essential for precision prevention. We aimed to assess whether objectively measured daily step count and Polygenic Risk Score (PRS) predict the incidence of T2D, and to evaluate the predictive performance of machine-learning models.

We conducted a prospective cohort study of 4,589 adults in the NIH All of Us Research Program with valid Fitbit step data and whole-genome-derived PRS, excluding individuals with T2D before or within a 180-day lead-in period. Incident T2D was defined by HbA1c ≥6.5%, plasma glucose ≥126 mg/dL, or an All of Us T2D condition record. We evaluated Cox and machine-learning survival models.

Over a median of 2.92 years (15,340 person-years), 265 participants developed T2D (5.77% cumulative incidence; 17.27 per 1,000 person-years). Risk-reducing thresholds were ∼7,000 steps/day (p<0.001) but differed by PRS group (∼7,800 high vs ∼5,800 low; p<0.001). Each additional 1,000 steps/day was associated with lower T2D risk (aHR 0.83, 95% CI 0.79-0.88; p<0.005), whereas each 1-SD higher PRS was associated with higher T2D risk (aHR 2.62, 2.32-2.96; p<0.005). Adding steps to a clinical model increased the C-index from 0.748 to 0.774; adding PRS increased it to 0.867. Penalized Cox achieved the highest discrimination (C-index 0.859), followed by survival SVM (0.85) and classical Cox (0.846). Calibration was best with Random Survival Forests.

Step-count thresholds for reducing T2D risk varied across genetic risk groups, indicating that step counts do not confer uniform protection. Step counts and PRS provided independent, complementary predictive information, and their combination improved prediction of incident T2D.