Design: Systematic review and Bayesian network meta-analysis of randomised trials
Population: Hospitalised non-ICU adults; 6 trials, 13,716 observations in the mortality network
Primary outcome: In-hospital or 30-day mortality; unplanned ICU transfer
Effect: Mortality OR vs standard care: continuous monitoring 0.70 (95% CrI 0.36–1.29); predictive alerts 1.22 (0.61–2.31)

This systematic review and Bayesian network meta-analysis compared three ways of catching deterioration on general wards — rule-based electronic alerts, predictive-model alerts and continuous physiological monitoring — against standard care. Of 28 trials, nine met strict criteria and six contributed to each main network (about 13,700 observations for mortality).

None clearly reduced mortality: odds ratios were 0.91 (95% credible interval 0.35–2.30) for rule-based alerts, 1.22 (0.61–2.31) for predictive-model alerts and 0.70 (0.36–1.29) for continuous monitoring. ICU transfer results were similar. Expanded rule-based response systems did not clearly reduce cardiac arrest (OR 0.94, 0.77–1.14), and length of stay did not change. Confidence in every comparison was rated very low.

Continuous monitoring leaned favourable but imprecisely. The authors' conclusion is the useful one: the effect of any system probably depends less on the algorithm than on who is monitored, how alerts reach clinicians and what response follows.

Do not assume a sophisticated predictive alert will outperform a well-run early warning score.
Put effort into the response pathway — who is called, how fast, and with what authority to escalate.
Target continuous monitoring at high-risk groups, such as post-operative patients on opioids, rather than every bed.
Audit alert fatigue on your ward: alerts that are ignored cannot help.
Treat vendor claims of mortality benefit with caution; randomised evidence does not yet support them.

Why it matters

It challenges the assumption that buying a smarter alert system will reduce ward deaths.

The statistics, in plain English

Every credible interval here crosses 1.0, and several are very wide (0.35 to 2.30), so the data are compatible with benefit, no effect or harm. 'Very low confidence' in the CINeMA framework means the true effect may be substantially different from the estimate.