Randomized Clinical Trials or Convenient Controls: TREWS or FALSE?

Shamim Nemati, Supreeth P. Shashikumar, Andre L. Holder, Gabriel Wardi, Robert L. OwensmedRxiv 2022 · DOI

Abstract

We read with interest the Adams et al. report of the TREWS machine learning (ML)-based sepsis early warning system. The authors conclude that large-scale randomized trials are needed to confirm their observations, but assert that their findings indicate the potential for the TREWS system to identify sepsis patients early and improve patient outcomes, including a significant decrease in mortality. However, this conclusion is based upon a comparison of those whose alert was confirmed vs. not confirmed within 3 hours, rather than random allocation to TREWS vs. no TREWS. Using data from over 650,000 patient encounters across two distinct healthcare systems, we show that the findings of Adams et al. are likely to be severely biased due to the failure to adjust for ‘processes of care’-related confounding factors.

Introduction: The Sepsis AI Evaluation Challenge

Sepsis remains a leading cause of hospital mortality, making early intervention a critical clinical goal. Machine learning (ML) models, such as the Targeted Real-time Early Warning System (TREWS), are designed to identify septic patients hours before traditional clinical recognition. However, evaluating these models in real-world settings is notoriously difficult.

The original TREWS study (Adams et al., Nature Medicine 2022) reported that patients whose alerts were confirmed by a clinician within three hours experienced significantly lower mortality than those whose alerts were confirmed later. This paper critiques the fundamental logic of that evaluation: the "control group" was not chosen at random. Instead, the study compared patients who received prompt care to those who did not. This introduces massive bias, as the speed of a clinician's response is often dictated by the patient's obvious severity or hospital workload—factors that independently influence survival.

Methods: The Coin-Flip Protocol

To demonstrate the impact of this bias, the authors employed causal inference techniques, specifically Directed Acyclic Graphs (DAGs), to map the relationships between patient data, clinician behavior, and outcomes. They designed a "Thought Experiment" called the Coin-flip Protocol to see if a random process could replicate the TREWS results.

Comparison of the standard process of care, the TREWS experiment, and the authors' coin-flip thought experiment.

The protocol involved:

  1. Random Alerts: Firing "alerts" based on a simple coin flip starting from the time of triage.
  2. Biased Assignment: Assigning patients to a "study arm" or "control arm" based on whether a clinician ordered a lactate test (a common sepsis marker) around the time of the alert. This mimics the "processes of care" bias suspected in the original study.
  3. Backdoor Adjustment: Using logistic regression to determine if the apparent mortality benefit vanished after adjusting for these care-related confounders.

The study utilized a massive dataset of over 650,000 patient encounters from two distinct healthcare systems: UC San Diego (Cohort A, 231,441 encounters) and Grady Hospital/Emory (Cohort B, 431,780 encounters).

Key Insights: The "Backdoor Path" to Spurious Success

The authors suggest that many sepsis AI models are not truly "predicting" sepsis; they are detecting that a clinician has already initiated treatment (e.g., by ordering labs). If an AI triggers because a doctor is already acting, it cannot take credit for the improved outcome.

In causal inference, a "backdoor path" is a common cause that creates a spurious correlation. Here, "clinical suspicion" leads to both an earlier alert (via more frequent data entry) and better outcomes (via faster treatment). Failing to "block" this path by adjusting for it makes the AI appear effective when it is merely a bystander to high-quality clinical care.

Key Results: Replicating the "Statistical Illusion"

The results of the Coin-flip Protocol were startlingly similar to the original TREWS findings, proving that a random process can appear to "save lives" if the data is grouped incorrectly.

Table showing patient characteristics and outcomes for Cohorts A and B, demonstrating the spurious mortality benefit in the coin-flip experiment.
  • Spurious Mortality Benefit: In Cohort A, the "study arm" (randomly alerted) showed a 3.3% mortality rate compared to 6.4% in the control arm ($p < 0.001$).
  • The Illusion of Success: Despite the coin flip appearing to reduce mortality by nearly 50%, the overall population mortality remained unchanged. The "benefit" was entirely an artifact of how patients were bucketed based on clinician behavior.
  • Adjustment Erases the Effect: Once a "backdoor adjustment" for the presence of lactate measurements was performed, the association between the alert and survival disappeared. The adjusted odds of mortality for the control group remained high (2.21 in Cohort A; 1.6 in Cohort B) even though the "intervention" was purely random.

Conclusion and Future Directions

While this study is a retrospective critique, it highlights the dangers of using "convenient controls" in observational studies of medical AI. The authors do not claim the TREWS model is useless, but rather that its reported benefits have not been rigorously proven.

To move forward, the authors strongly advocate for:

  • Prospective Randomized Clinical Trials (RCTs): These remain the gold standard for validating sepsis early warning systems.
  • Causal Modeling: Future observational studies of ML models should include DAGs to explicitly state and adjust for assumptions regarding data generation in clinical environments.

Summary updated

Abstract

We read with interest the Adams et al. 1 report of the TREWS machine learning (ML)-based sepsis early warning system. The authors conclude that large-scale randomized trials are needed to confirm their observations, but assert that their findings indicate the potential for the TREWS system to identify sepsis patients early and improve patient outcomes, including a significant decrease in mortality. However, this conclusion is based upon a comparison of those whose alert was confirmed vs. not confirmed within 3 hours, rather than random allocation to TREWS vs. no TREWS. Using data from over 650,000 patient encounters across two distinct healthcare systems, we show that the findings of Adams et al. are likely to be severely biased due to the failure to adjust for ‘processes of care’-related confounding factors.
BibTeX
@misc{nemati2022randomized,
  author = {Shamim Nemati and Supreeth P. Shashikumar and Andre L. Holder and Gabriel Wardi and Robert L. Owens},
  title = {{Randomized Clinical Trials or Convenient Controls: TREWS or FALSE?}},
  year = {2022},
  doi = {10.1101/2022.08.08.22278526},
  howpublished = {medRxiv},
}
Figure