Abstract / Summary
Abstract Background . Dynamic, gestation-anchored risk prediction could give clinicians lead time to act on pregnancy loss and stillbirth. Administrative claims databases are attractive substrates for such models because of their scale and population coverage, but whether claims support above-chance dynamic prediction of these outcomes at clinically actionable gestational time points is unestablished. Methods . Using the Virginia All-Payer Claims Database (2018--2023; 344,560 anchored pregnancy episodes), we built a discrete-time competing-risks survival framework distinguishing live birth, early pregnancy loss ($<$20 weeks), and stillbirth ($\geq$20 weeks), excluding induced termination. A gestational clock was anchored via ICD-10-CM Z3A codes; episodes were expanded to person-periods; predictions were made at gestational landmarks with a two-week gap/blanking window to prevent peri-event leakage. An evidence-based, code-level feature set was compared against a broad 3-character-prefix set using penalized discrete-time models, and evaluated with competing-risks cumulative incidence, time-dependent AUC, IPCW Brier / IPA, and calibration, under patient-level random and temporal splits. Results . Three findings emerged. Naive weekly-hazard models were severely inflated by same-week coding contamination: early-loss AUC fell from $0.833$ to $0.61$ once a two-week gap was enforced, with the events in the cell dropping from $378$ to $48$. An apparent ''worse-than-nothing'' index of prediction accuracy proved to be a two-part artifact (a per-period-versus-cumulative scale mismatch together with rare-cause magnitude inflation from balanced class weights), and resolved fully under proper competing-risks cumulative-incidence computation and recalibration (corrected IPA $\approx 0$). With contamination removed and evidence-based features properly constructed, neither outcome was forecastable above chance at actionable landmarks: stillbirth time-dependent AUC peaked at $0.56$ with a flat discrimination curve across gestation, and broad granularity did not help ($0.52$). A marker of recognized high-risk surveillance (ICD-10-CM O09) was \emph{anti}-predictive, consistent with effective surveillance biasing toward live outcomes. Conclusions . In this APCD extract, linear competing-risks models on properly constructed claims features did not predict early loss or stillbirth above chance at clinically actionable gestational time points. The result delineates a ''claims floor'' that contrasts with the biophysical ceiling of ultrasound/Doppler-based models, because the antecedents that carry predictive signal are coded at the event rather than antenatally. Beyond the substantive null, we provide two reusable methodological cautions for claims-based dynamic prediction.