Moving From In Silico to In Clinico Evaluations of Machine Learning-Based Interventions in Critical Care*

The patient at risk for critical illness seems a likely beneficiary of machine learning (ML)-based interventions. Their care decisions are high stakes, complex, replete with uncertainty, and informed by rapidly changing, high volume, and data inputs. Because ML methods are so suited to clinical decision support (CDS) for critical care, deterioration models have been among the earliest (1,2) and most popular (3) targets of ML model development efforts. But do we know how ML models affect care decisions or patient outcomes? The majority of evidence for ML models in critical care is in silico or based on a model's ability to predict an outcome in a retrospectively collected dataset. What the field largely lacks is robust evidence about a model's impact in clinico or how a model might affect care processes and patient outcomes when deployed into clinical workflows. Evidence about safety and equity are also needed but largely absent. Deterioration alerts are not entirely benign as they allocate resources toward some patients and away from others (4). In this issue of Critical Care Medicine, Levin et al (5) report the results of a pragmatic, nonrandomized, open-label, cluster trial (Realtime Streaming Clinical Use Engine for Medical Escalation [ReSCUE-ME]) of a ML-based early warning system to predict in-hospital deterioration among ward patients at a large academic hospital in New York City. In doing so, the authors make an important contribution to the literature by providing in clinico evidence for how this model affects both treatments and outcomes. STUDY DESCRIPTION The ML model, named Modified Early Warning Score (MEWS)++, is a random forest trained on laboratory tests, vital signs, electrocardiogram data, and structured nursing documentation to predict the probability of a composite outcome of transfer to an ICU or in-hospital death. Before initiating the ReSCUE-ME trial, the authors reported promising performance of MEWS++ in an in silico evaluation (6). The study team deployed the ML model in a general medicine and in a general surgery ward (intervention). Another pair of general medicine and surgery wards served as controls. Each patient's initial ward admission location determined their group assignment. Before the trial, a silent test of the model identified a primary trigger threshold that would produce an estimated eight alerts per day to the nursing team via a voice over IP device. A secondary alert was linked to a trigger threshold expected to prompt six rapid response team (RRT) activations each day. Because the model updated its predictions any time new clinical data became available, the team instituted several filtering criteria to minimize alarm fatigue and reduce the burden on clinicians from potentially redundant alerts. In the two control wards, the ML model generated predicted probabilities in the background but did not send any alerts. The original trial protocol described the primary outcome as a composite of the rate of care escalation and RRT initiations per thousand patient bed days. This outcome was modified to include only the escalation rate because of their causal interrelatedness. The sample size of 6426 participants was estimated based on an expected 20% relative reduction in escalations from a historical escalation rate of 10% to 8%. The analysis followed the intention-to-treat principle based on the initial group assignment and was restricted to each eligible patient's first hospital admission. Because of the nonrandomized design, an inverse probability of treatment weight (IPTW) was used to account for differences in observed baseline patient characteristics. This adjustment was important because a bariatric and a hepatology service were in the intervention and control wards, respectively. These ward characteristics influenced patient selection to the study groups in systematic ways. STUDY FINDINGS In an unadjusted analysis, there was no difference (absolute difference, 0.96%; 95% CI, –2.75% to 4.67%) in escalation between the intervention and control wards. In the IPTW analysis, patients in the intervention group had substantially higher rates of escalation to ICU care per 1000 patient bed days (incidence rate ratio, 1.43; 95% CI, 1.16–1.78). This latter result was robust to a sensitivity analysis using propensity score matching. Patients in the intervention wards were more likely to receive a medication order and electrocardiogram within 6 hours of an alert firing compared with when an alert occurred but was not delivered in the control wards. The increase in medication orders was primarily driven by cardiovascular support (defined as either "autonomic, cardiac, cardiovascular, or diuretic medications, or IV fluid therapy"). Finally, the composite mortality outcome was also lower in the group of patients who received an alert with an absolute risk difference –2.4% (95% CI, –4.5% to –0.2%). These secondary findings were also consistent in the IPTW and propensity score analyses. In summary, the study reported that deploying this ML model resulted in no difference in ICU transfers in the primary analysis, but a difference was observed in the weighted analysis that accounted for patient characteristics. The ML-driven alerts were associated with more rapid ordering of cardiovascular medications and IV fluids and a significant mortality reduction. DISCUSSION These promising findings add to a sparse but growing body of experimental literature on the beneficial effects of ML models to prompt early interventions forward patients at risk of deterioration. The study by Levin et al (5) has several important strengths and highlights key lessons for establishing a strong causal link between the use of ML deterioration models and clinical outcomes (7). First, the trial was terminated early in March 2020 due to the COVID19 pandemic. At the time of termination, the investigators had enrolled 2740 patients with complete data or 43% of the planned number of participants. Early trial termination due to the COVID19 pandemic is not uncommon, affecting at least several thousand studies, and raises both statistical and ethical questions about how to approach the incomplete data accrued in a terminated study (8). Following termination, the authors thoughtfully estimated a futility index in a post hoc analysis to inform their decision to ultimately report the study's results. Early termination, even without a pandemic, is a common threat to ML-based trials and can be avoided when incentives for clinical investigators and hospital leaders are closely aligned (9). Second, the study used a nonrandomized design, making causal effect estimates susceptible to confounding through selection. Among the rigorous strategies employed to account for confounding, an IPTW analysis adjusted for observed confounders, such as chronic comorbidities associated with admission to the study wards with different clinical subspecialties (10). Indeed, after weighting, the study showed very good balance between the intervention and control populations. The study goes further and reported the results of a propensity matching study, an alternative casual strategy to account for observed confounders (11), and finds consistent results. But ML-based deterioration alerts are ideally studied in a randomized design because of important causal threats from confounding by unobserved factors (12). In the study by Levin et al (5), incompletely observed characteristics of patients and wards likely affected both the treatment assignment and the outcomes. This bias alters the effect estimates in uncertain ways in the absence of a randomized design and limits the causal link between the intervention and clinical outcomes. Third, how should hospitals interpret the results of the study by Levin et al (5) when deciding whether to deploy MEWS++ for their own patients? The MEWS++ model is now used at the authors' home institution to generate a prioritized list of patients at high risk for deterioration in nine stepdown units (author communication). A dedicated team of intensivists rounds each day on the 10–15 sickest patients whose score is above a predetermined threshold. This modification to the studied protocol was chosen for its sustainability and scalability. Hospital leaders face challenging scientific, operational, and regulatory decisions about adopting ML models. Even among leading academic medical centers, the approaches to governing such systems are still evolving and lack consensus (13). The Food and Drug Administration (FDA) recently classified nearly all predictive CDS software for critical care as a device, thereby requiring a formal regulatory review and authorization. But a recent study of FDA authorizations for critical care CDS found no examples of in clinico studies or equity evaluations considered during the review process (14). Thus, the scientific limitations of studies like this one are foregrounded by broader regulatory and ethical questions about the oversight of critical care focused CDS devices like MEWS++. Finally, heterogeneity in deterioration definitions appropriately reflects variation in hospital level priorities but also limits systematic model comparisons (7). A recent expert survey identified a consensus definition of an adult inpatient decompensation event (15) but has not yet been widely implemented. These differences parallel broader tensions in ML model development between local needs and generalizable knowledge. In summary, too little is known about what happens when a ML model is deployed in hospitals wards to detect clinical deterioration. Technical, operational, and regulatory uncertainty overshadow the deluge of published prediction models for this purpose, most of which have never prompted an alert to a real clinician. But the study by Levin et al (5), in deploying and evaluating the effects of such a model on clinical care and patient outcomes, is another important step in moving the evidence for clinical deterioration ML models from byte to bedside.