prognostication                 

Accurately predicting neurological outcomes after cardiac arrest remains one of the greatest challenges in resuscitation science.

A central obstacle is the potential bias introduced by withdrawal of life-sustaining therapies (WLST). Early withdrawal can create a self-fulfilling prophecy, where patients who otherwise would recover instead die after WLST.  We have shown WLST is common across care settings, often affecting even those who experts think might otherwise have recovered, thereby contributing to avoidable deaths. This can bias the results of clinical trials, prognostic models, and clinical decision making.

Building on this foundation, we have advanced approaches that combine neurophysiology via longitudinal EEG analysis, imaging, and clinical factors to improve early risk stratification and prognostic accuracy while minimizing bias. We developed frameworks that integrate multimodal phenotyping with time-varying/Bayesian models that explicitly account for competing risks such as awakening, WLST, and death, enabling dynamic updating of predictions during ICU care. This work is supported by NIH funding and related projects aimed at reducing bias and providing actionable prognostic information for clinicians and families.

We have also advanced advanced analytical strategies for EEG interpretation, including group-based trajectory modeling (GBTM), machine-learning classifiers for continuous waveforms, and Bayesian dynamic models that adapt forecasts as new information accrues. These tools not only improve accuracy, but also open the door to decision-support systems that reflect how clinicians and families process prognostic information. Parallel work in decision science emphasizes how cognitive factors, uncertainty, and communication shape prognostic decision-making at the bedside. Together, these efforts aim to create a more reliable, patient-centered foundation for post-arrest prognostication—reducing avoidable bias and supporting recovery whenever possible.

Recent Work

  • We analyzed observational cohorts to quantify how withdrawal of life sustaining therapy for perceived poor neurological prognosis (WLST N) can bias outcome estimates and showed that censoring outcomes after WLST N could yield more reliable awakening predictions than treating non awakening post WLST as ground truth.

  • We examined how much early prognostic information is needed for highly specific poor outcome prediction and argued for parsimonious, interpretable models that incorporate EEG. 

  • A narrative review from our group synthesizes ethical frameworks, timing, and safeguards for WLST after cardiac arrest, emphasizing strategies to reduce self fulfilling prophecy and promote equity in decision making. 

  • We demonstrated that duration of coma predicts short-term functional but not long-term survival, showing that even patients with prolonged coma can enjoy excellent post-acute survival.

  • Through qualitative interviews, we defined the mental model used by clinician to acquire and interpret prognostic data under uncertainty, revealing cognitive pitfalls and opportunities for structured improvements. 

  • Our national cross sectional analysis revealed a paucity of neuroprognostic testing after arrest, highlighting practice variation and missed opportunities to reduce uncertainty and bias in outcome prediction.