Background: An AI-enabled early-alert system embedded in routine clinical management platforms may help healthcare providers identify patients with head and neck cancer (HNC) at risk of experiencing an adverse outcome, especially an imminent catastrophic health event. Developing such AI systems requires 1) identifying HNC-meaningful catastrophic health events, focusing on those that are detrimental yet potentially preventable, and 2) reviewing data sources and methods used to capture these events to ensure accuracy and replicability. This is an ongoing review.
\n\nMethods: In this PRISMA-guided scoping review, we worked with a medical librarian to conduct a comprehensive search using MEDLINE, Embase, CINAHL Plus, APA PsycINFO, and Cochrane CENTRAL databases for peer-reviewed, English-language journal articles published between 2015–2025. Non-trial studies documenting health outcomes of adult HNC patients labelled as “catastrophic” or a similar descriptor were included. Two reviewers independently screened all studies and extracted the data in duplicate. Any common terminology criteria for adverse events (such as the CTCAE) were excluded as we aimed to capture unstandardized outcomes with detrimental impact on HNC patients. This abstract presents the results of studies conducted in North America between 2020–2025.
\n\nResults: Our search yielded 7,025 unique citations. After title/abstract screening, 758 studies were eligible for full-text review; 143 were included in the overall study, with 16 meeting eligibility for this analysis. Of the 16 studies, 15 were conducted in the USA and 1 in Canada, representing 241,147 adults with mainly oropharyngeal or oral cavity cancer undergoing curative treatment. Three domains of catastrophic health events were reported: carotid blowout, airway compromise and other severe surgical complications (n=7, 43.8%); sudden/premature deaths (n=8, 50.0%); and other events (n=16, 100%) including clinically diagnosed depression (n=1), patient-reported severe pain/fatigue (n=2) and other symptoms from the MD Anderson Symptom Inventory (n=1), and acute care encounters including unplanned hospital admissions (n=1), readmissions (n=11) and ICU admissions (n=1). Of the 7 studies reporting severe surgical complications, 3 used electronic medical records (EMR) to capture these events, while the remaining studies used health administrative data (i.e., hospitalization records with the corresponding ICD diagnostic code) or surgical quality assurance data (e.g., the National Surgical Quality Improvement Program or NSQIP). Meanwhile, all studies documenting sudden/premature deaths or acute care encounters used routinely collected health administrative data to determine these events. Of the 3 studies reporting severe patient-reported symptoms, 2 used the EMR, and 1 used a patient questionnaire.
\n\nConclusions: Outcomes that can be determined with certainty by using routinely collected health administrative or registry databases, such as all-cause mortality and unplanned hospitalizations, are a realistic starting point for AI-based early-alert system development in HNC. Outcomes that require more clinical expertise, especially those that require patient interviews/chart reviews of EMR, may require considerably more effort to define before using them to develop a prediction algorithm. The preliminary results of this review suggest that catastrophic health events in HNC are captured through relatively diverse data sources. These findings have actionable implications for designing clinically relevant and feasible AI-based early-alert systems for catastrophic health events in HNC.