Patient factors associated with survival after critical care interhospital transfer

Andrew P. Reimer, Wei Dai, Nicholas K. Schiltz, Jiayang Sun, Siran M. KoroukianFrontiers in Disaster and Emergency Medicine 2024 · DOI

TL;DR

Researchers developed a machine learning model using structured electronic health record (EHR) data to predict mortality for patients undergoing critical care interhospital transfer. The study found that while patient physiology (labs and vitals) strongly predicts survival, the mode of transport—helicopter versus ground ambulance—was not a significant predictor of mortality for the general population.

What Problem Does This Paper Address?

Interhospital transfer (IHT) is a high-stakes component of modern healthcare, with over 1.6 million patients moved between facilities annually in the U.S. alone. Clinicians must decide not only if a patient should be moved, but how (ground vs. air) and when.

While "time-sensitive" cases like major trauma or strokes clearly benefit from rapid air transport, nearly 70% of transfers involve non-time-sensitive but critically ill patients (e.g., sepsis or multi-organ failure). For these patients, helicopter transport is 3 to 20 times more expensive than ground transport, yet there is little evidence it improves outcomes. This paper seeks to identify the specific patient factors that actually drive survival to build better clinical decision-support tools.

Methods

The authors employed a "multi-method" analytical approach to handle the high dimensionality and sparsity of EHR data:

  • Variable Reduction: They used Random Forest (RF) to sift through over 2,700 unique diagnosis codes and laboratory parameters, identifying the most impactful predictors.
  • Predictive Modeling: They compared Logistic Regression (LR) for interpretability and Gradient Boosting for predictive power. They built four nested models, starting with demographics and incrementally adding laboratory results and vital signs.
  • Relationship Mapping: Multiple Correspondence Analysis (MCA) was used to visualize the "proximity" between patient variables and outcomes (death vs. survival).
  • Benchmarking: The final model's performance was compared against the APACHE IV score, a gold-standard clinical tool for predicting ICU mortality.

Datasets

The study utilized the Medical Transport Data Repository, a retrospective dataset from a large health system in Northeast Ohio (a main campus plus 11 affiliated centers).

  • Sample Size: 11,497 adult patients (aged 21+) who underwent critical care transfer between 2010 and 2017.
  • Data Types: Structured EHR data including diagnoses, laboratory results (albumin, BUN, glucose, etc.), vital signs (BP, pulse, temp), and social history.

Key Insights and Ideas

  • Physiology Trumps Logistics: The most powerful predictors of mortality were age, ventilator usage, and specific biomarkers (Albumin, INR, Platelets, BUN). Interestingly, the "mode of transport" (helicopter vs. ground) did not significantly impact survival in the general population models.
  • The "Structured Data" Advantage: By using only structured EHR data (data already in fixed fields), the researchers achieved an AUROC of 0.85. This is significant because it matches the performance of the APACHE IV score without requiring the manual, labor-intensive data entry that APACHE often demands.
  • The Helicopter Paradox: While helicopters are often defaulted to for the "highest level of care," the data suggests that for many patients, the clinical expertise of the transport team and the stability of the patient are more critical than the speed of the aircraft.
  • Survival Indicators: Factors like comorbid hypertension and being transferred from an Emergency Department (rather than an ICU) were actually associated with increased odds of survival, likely reflecting a "healthier" subset of the critical care population.

Key Results

  • Model Performance: The final Gradient Boosting model (Model 4) achieved an AUROC of 0.85 and a testing accuracy of 78%.
  • Mortality Rate: The overall mortality rate (including discharge to hospice) was 32.9%.
  • Top Predictors: Age was the most important variable (Relative Variable Performance = 100), followed by Albumin (88.25) and Ventilator usage (67.94).
  • Transport Mode: In the logistic regression model, helicopter transport had an Odds Ratio (OR) of 1.14 but was not statistically significant (p > 0.05), suggesting it isn't a primary driver of survival for the broad cohort.

Limitations and Future Work

The authors acknowledge several limitations:

  • Selection Bias: The data comes from a single health system, which may have specific transport protocols that don't generalize everywhere.
  • Natural Randomization: Factors like weather or helipad availability often dictate transport mode regardless of clinical need, which might "muddy" the data regarding the efficacy of air transport.
  • The "Time-Sensitive" Gap: Future research is needed to identify the specific sub-phenotypes of patients who aren't currently triaged as "time-sensitive" but would still benefit significantly from the speed of a helicopter.

The ultimate goal is to embed these models into EHR systems to provide real-time "transport triage" advice to bedside clinicians.

Summary updated

Abstract

Objective To identify the factors that predict mortality post-transfer and develop a comprehensive mortality prediction model capable of supporting pre-transfer decision making. Materials and methods Electronic health record data from the Medical Transport Data Repository of a large health system hospital in Northeast Ohio that consists of a main campus and 11 affiliated medical centers. We retrospectively analyzed patient data from the referring hospital encounter prior to interhospital transfer. All patient data including diagnoses, laboratory results, medication, and medical and social history were analyzed to predict in-hospital mortality post-transfer. We employed a multi-method approach including logistic regression, gradient boosting, and multiple correspondence analysis to identify significant predictors of mortality as well as variables that are clinically useful to inform clinical decision support development. We identified all patients aged 21 and older that underwent critical care transfer in the health system between 2010 and 2017. Results We found that age, laboratory results (albumin, INR, platelets, BUN, leukocyte, hemoglobin, glucose), vital signs (temperature, respirations, pulse, systolic blood pressure, pulse oximetry), and ventilator usage are the most predictive variables of post-interhospital transfer mortality. Using structured data from the EHR we achieved the same performance as APACHE IV within our health system (0.85 vs. 0.85). Lastly, mode of transport alone was not a significant predictor for the general population in any of the outcome models. Conclusions Our findings provide a foundation for the development of decision support tools to guide transport referrals and identified the need for further inquiry to discern the role of mode of transport to enable future inclusion in decision support approaches. Further inquiry is needed to identify factors that differentiate patients not triaged as time-sensitive transfers but still require helicopter intervention to maintain or improve post-interhospital transfer morbidity and mortality.
BibTeX
@inproceedings{reimer2024patient,
  author = {Andrew P. Reimer and Wei Dai and Nicholas K. Schiltz and Jiayang Sun and Siran M. Koroukian},
  title = {{Patient factors associated with survival after critical care interhospital transfer}},
  booktitle = {Frontiers in Disaster and Emergency Medicine},
  year = {2024},
  doi = {10.3389/femer.2023.1339798},
}
Figure