AI Model Predicts In-Hospital Cardiac Arrest After TSCI

Key Takeaways

• Among adults admitted primarily for acute TSCI to US and Chinese ICUs, first-day measures distinguished later documented cardiac arrests in independent external testing.
• Approximately three-quarters of subsequent arrests in the Chinese cohort fell within the highest-risk tenth at the locked threshold.
• Potassium, APACHE II, bicarbonate, temperature, and mean arterial pressure contributed most to predicted risk in SHAP analysis, without establishing causal effects.
• Favorable modeled decision-curve results did not establish that using the model reduces arrests or improves survival.

Acute traumatic spinal cord injury (TSCI) can disrupt autonomic control and destabilize physiology during intensive care unit (ICU) treatment. In-hospital cardiac arrest (IHCA) may be difficult to anticipate before overt deterioration, leaving early risk identification unresolved. Could routinely recorded findings from the first ICU day distinguish patients who subsequently arrest from those who do not?

In a traumatic spinal cord injury prediction study, investigators retrospectively developed an IHCA model among adults admitted primarily for acute TSCI. US records came from Medical Information Mart for Intensive Care IV (MIMIC-IV) and the electronic ICU Collaborative Research Database (eICU-CRD). The development cohort included 996 patients, with 57 subsequently documented arrests. Overt hemodynamic deterioration typically prompts a response, but the earlier identification of arrest risk remains unresolved. First-day measurements preceded ascertainment of IHCA, which continued through day 7 or earlier discharge or death.

Independent testing used records from six Chinese hospitals, with 558 patients and 51 arrests. Eligible predictors were recorded during the first ICU day; arrests before or during that window were excluded. An event required time-stamped documentation of abrupt loss of cardiac mechanical activity or an unequivocal equivalent. US exclusions included preprediction do-not-intubate (DNI) or combined do-not-resuscitate/do-not-intubate (DNR/DNI) orders. The use of US databases for development and Chinese hospitals for testing makes these findings relevant to US care, but the model has not been prospectively adopted there.

Investigators used observations from the first 24 ICU hours and selected ten inputs: spinal injury level, paralysis type, injury mechanism, vasopressor use, Acute Physiology and Chronic Health Evaluation II (APACHE II) score, serum bicarbonate, sodium, potassium, mean arterial pressure, and body temperature. Imputation and feature selection used training data only. The ensemble assigned predicted risk of subsequent IHCA, not an autonomous treatment decision.

Univariable screening, Boruta, and group lasso fixed the inputs before 11 base learners underwent fivefold cross-validation. Their out-of-fold probabilities fed an L2-regularized logistic stacking model. Training-fitted missForest imputation was applied to internal validation without refitting. The predictors, ensemble, and training-selected classification threshold were locked before Chinese testing against APACHE II, Sequential Organ Failure Assessment (SOFA), and training-derived logistic regression. Evaluations covered discrimination, calibration, modeled net benefit, lift, and Shapley additive explanations (SHAP), not outcomes under an alert intervention.

In external validation, the ensemble achieved an area under the receiver operating characteristic curve (ROC-AUC) of 0.894 and precision–recall average precision (PR-AP) of 0.774. These measure prediction, not reductions in arrests. Training and internal validation showed higher discrimination and precision–recall performance. In the Chinese cohort, the ensemble had stronger discrimination, precision–recall performance, and a more balanced classification profile than APACHE II, SOFA, or the unchanged training-derived logistic regression.

The authors described calibration as acceptable; intervals for external calibration-in-the-large and slope included their respective ideal values. Decision curves suggested model-estimated net benefit across several thresholds, rather than observed benefit from acting on predictions. Global SHAP rankings reflected how biochemical, severity, temperature, and perfusion inputs contributed to computed risk; they did not identify independently proven mechanisms of arrest. The reported injury-level SHAP rankings came from the training set and were not separately validated in the external cohort.

Retrospective prediction cannot establish whether model-guided alerts change arrest rates, survival, or alarm burden. The training set contained 43 arrest events, and feature selection ran once on the full training set rather than within each cross-validation fold, allowing optimism in training out-of-fold estimates; neither locked validation cohort contributed to fitting. US–China differences in case mix, potassium and mean arterial pressure distributions, recording density, and stabilization before Chinese ICU transfer complicate transportability. An interactive online calculator implements the model, but its availability does not demonstrate clinical benefit. Alert fatigue remains a proposed concern, not a measured result.

Wang and colleagues called for prospective evaluation before routine implementation, including whether use of the locked model changes clinical outcomes. Neither external predictive performance nor modeled decision-curve benefit establishes an effect on arrests, survival, or alert burden in practice.

Clinician Questions

Did a DNR order exclude ICU patients with traumatic spinal cord injury from cardiac-arrest risk prediction?

In the US TSCI ICU development cohort, an isolated DNR order did not exclude a patient: the target was documented cardiac arrest regardless of whether resuscitation was attempted. Preprediction DNI or combined DNR/DNI orders did exclude US patients; the Chinese sites did not use those orders during the study period.

Were missing first-day ICU predictors imputed in the Chinese traumatic spinal cord injury validation cohort?

All first-day predictor fields were complete in the Chinese TSCI ICU validation cohort, so external values needed no imputation. Completeness in that cohort does not establish what data would be available at other hospitals using the model.

Did the cardiac-arrest prediction model rely on the same features across traumatic spinal cord injury levels?

In SHAP analyses of ICU patients with TSCI, potassium ranked first for upper cervical, other documented, and uncertain injury levels. Other prominent contributions to modeled arrest risk varied: injury mechanism and APACHE II for upper cervical injury; APACHE II, temperature, and bicarbonate for other documented levels; and injury mechanism and paralysis type when level was uncertain. These rankings do not establish causal effects or differences in clinical outcomes.