Sequential monitoring of risk-adjusted mortality in intensive care commonly uses control-chart approaches such as the risk-adjusted CUSUM and EWMA. These methods generate alarms that indicate potential deterioration but do not directly provide a calibrated, continuously updated estimate of the standardized mortality ratio (SMR) together with quantified uncertainty. The need addressed here is a monitoring approach that simultaneously produces timely alarms and an interpretable, uncertainty-aware SMR trajectory for clinicians and registry stewards.
STREAM-SMR is a conjugate gamma–Poisson dynamic generalized linear state-space model designed to produce filtered estimates of the SMR while supporting sequential monitoring. In the model the latent log-SMR evolves through a discount factor delta that controls temporal smoothing. The method yields closed-form updates and is exact for months with zero observed deaths. The alarm statistic used for monitoring is the posterior exceedance probability P(SMR > 1), which provides a direct Bayesian probability that observed mortality exceeds expected mortality at each time point.
To evaluate performance, the author compared STREAM-SMR with the risk-adjusted CUSUM and risk-adjusted EWMA using a Monte Carlo simulation study. STREAM-SMR was run with three discount factors: delta = 0.90, 0.95, and 0.97. Comparisons were stratified by facility volume (50, 200, and 800 annual admissions) and by five change scenarios, including sustained steps, gradual drift, transient deterioration, and improvement. All three monitoring methods were calibrated via Monte Carlo so that each achieved a common nominal 5% false-alarm probability over a 60-month surveillance horizon. Each cell in the factorial simulation had 1,000 replications. Calibration used a protocol frozen before evaluation and aggregated registry reports to set scenario parameters.
Across sustained-shift scenarios and with delta = 0.90, STREAM-SMR closely matched the detection frontier of the theoretically optimal risk-adjusted CUSUM, typically within one to two months of median detection delay. For example, when the SMR stepped to 1.5, median delays (STREAM-SMR vs CUSUM) were: large facilities (800 admissions) 6 versus 5 months, medium (200) 14 versus 14 months, and small (50) 20 versus 22 months.
Filtered SMR estimates returned by STREAM-SMR included 95% credible intervals whose empirical coverage was at least 91% in every scenario–stratum cell evaluated. Empirical false-alarm probabilities across all methods were reported to be close to the nominal 5% target, with observed values ranging from 0.035 to 0.066 in the simulations.
In the case of a three-month transient deterioration, STREAM-SMR detected changes faster conditional on detection. However, when considering overall detection probability (the chance the method signals at all for the transient event), the CUSUM had a higher overall detection probability in large facilities.
The author frames STREAM-SMR as a unified object for both monitoring and estimation. The discount factor delta is an explicit tuning parameter that sets the trade-off between smoothness of the estimated SMR trajectory and speed of detection. Lower delta (for example, 0.90) increases sensitivity to recent changes and yields detection speed closer to that of the CUSUM at the cost of less smoothing; higher delta yields smoother SMR estimates but slower detection. The reported performance suggests that, for a small detection-delay premium (one to two months relative to the CUSUM), STREAM-SMR provides a continuously updated, interpretable SMR with uncertainty quantification at every time point.
STREAM-SMR is described as computationally trivial at registry scale because the conjugate construction allows closed-form sequential updates and exact handling of zero-death months. The method’s outputs—filtered SMR estimates with credible intervals and a posterior exceedance probability alarm—are intended to be readily interpretable by clinicians and registry stewards, unlike standalone alarm charts that do not report an evolving SMR trajectory.
All simulation code, the frozen protocol, configuration, and seeds are openly archived at Zenodo (doi: 10.5281/zenodo.21991702). No patient-level data were used in the study; facility strata were calibrated to published aggregate registry reports only. The paper notes that the protocol was fixed before running evaluations.
The preprint declares that the author, K.O., is Representative Director of Jinen Co., Ltd. and serves as engineering staff for a non-profit organization operating a national ICU registry. No specific funding was reported for the study. The manuscript is a medRxiv preprint and has not been peer reviewed; the author confirms adherence to ethical reporting and that necessary approvals and consents were obtained as described in the source.