Principal FindingsOverviewThis feasibility study explored the use of LLMs to summarize phone calls between referring and accepting physicians in the setting of emergency care for patients with suspected STEMI and evaluated this process using standardized assessment tools. With the caveat that mean scores were derived from ratings with only slight agreement, the high (adapted) PDQI score suggests that the LLM-based approach yielded summaries that were accurate and reproducible. Transfer call summaries, particularly when compared to complete opacity, may represent a transformative solution to address data loss during interhospital transfers. This work demonstrates the feasibility of using an LLM to summarize transcripts of STEMI transfer calls and provides a foundation for further investigation of automated summarization of call audio in patient transfers.
Call Summary QualityNo PDQI domain received ratings under 4/5 when averaged over the summaries. The “thorough” and “hallucination free” domains had the lowest scores of the 8 measured domains. Whether a lower score in the “thorough” domain represents the LLM system behaving as designed given inherent limitations of how thorough a summary can be or reflects true information loss is unclear as this was not addressed in the comments left by raters. The distinction is important as redesigning the prompt could generate more thorough, albeit potentially longer summaries. However, the consistently high ratings for the “useful” domain are reassuring and suggest that the lower scores in the “thorough” domain are less likely to be due to diminished transfer of high-yield information.
We found that there were 2 types of hallucinations referred to in the comments regarding the “hallucination free” domain. The first was fabrication, which could clearly cause harm if patients’ management was changed due to inaccurate information. The second was clinical inference (such as inferring hemodynamic instability from words such as “pale” or “diaphoretic”), which was considered reasonable and valid by reviewers but was still labeled as a hallucination because the information was not explicitly verifiable via the call audio. In this initial study, the “hallucination free” rating could represent the worst-case scenario or upper bound, but future work might benefit from parsing further the definition of the hallucination domain by including fabrication and valid inference as 2 separate domains.
Although relatively uncommon, fabrications could potentially adversely influence patient care, particularly if there are no associated primary data included in the summary from which receiving physicians can draw their own conclusions (ie, a receiving physician may interpret vital signs differently from the LLM, but there is no analogous corroborating evidence to support whether a medication was given). Furthermore, if the LLM reported a categorical assessment of a variable such as vital signs and the primary data were not included in the summary, a miscategorized vital sign may prompt misclassification. However, there are strategies for mitigation of LLM hallucinations that were not within the scope of this project [17].
That the LLM had difficulty with the concept of time since the chest pain started is not surprising. Although LLMs have recently achieved significant advances in mathematical reasoning, this remains a challenging frontier in AI research [18]. The prompt we designed requested the number of minutes since the chest pain started, which may require knowledge of the time of the call, the time of onset of chest pain, and the ability to reason mathematically. An alteration of the prompt to request any relevant information from the call on the time when symptoms started as opposed to “the number of minutes since chest pain started” might improve the summary’s accuracy with respect to the timing of relevant events.
There are several additional issues that are important to consider prior to implementation. The average WER was 1.6% (SD 0.4%) for 6 randomly selected transcripts, which is reassuring; however, we did not thoroughly assess the conceptual accuracy of all transcripts, nor did we analyze common errors of transcription. Propagated errors of transcription may have degraded the quality of the summary. Assurance that the transcription model includes medical terminology and drug names is important, as is thorough assessment of transcription quality. Further prompt engineering may reduce the incidence of hallucinations, particularly with respect to the timing of events, including the onset of chest pain. For example, updating the prompt to request only timing that was explicitly discussed in the call may be helpful. Finally, to engender confidence, it may be reasonable to request quotations from the transcript so that physicians can look back to the primary data to confirm critical information.
LimitationsThere were several limitations to this study. The transcription software and LLM were chosen for convenience by the study team, and alternative software or alternate versions could produce different results. The sample size was small, which could limit overall generalizability, although the patient population was relatively homogenous (ie, patients being transferred to a tertiary care center due to concern about STEMI). The selection process included information-rich calls that contained complete clinical and decision-making information, which introduced selection bias. Performance on a randomly selected sample of calls would improve the generalizability of the study to the broader clinical environment, although using an enriched dataset for this feasibility study is a reasonable initial step.
Lack of a comparator such as human expert–generated summaries limits interpretability of the summary quality scores, and the study’s retrospective design limits our ability to understand the clinical adequacy of the summaries and downstream effects of summarization on patient care. A prospective study measuring the impact of incorporation of summaries into clinical workflow on patient outcomes (such as morbidity, mortality, time in the intensive care unit, and overall length of stay) compared to an expert-generated comparator could address some of these design limitations.
Although we did not test different prompts, the prompt used in this study was engineered to optimize the inclusion of key objective data while generating a summary that could be consumed efficiently by a busy clinician. Future work could address this limitation by having receiving physicians evaluate the overall utility of summaries generated using different prompts.
The second- and third-year internal medicine residents reviewing the summaries did not evaluate them from the perspective of interventional or critical care cardiologists; however, they had enough exposure to CICU and general cardiology rotations in their training to reasonably make assessments of summary quality. Furthermore, they are often the intended beneficiaries of call summaries (ie, the first-line recipients of post-PCI patients with STEMI). That said, their relative inexperience could lead to different assessments of the importance of certain elements of the call from those of a more experienced clinician, particularly for the “useful” and “thorough” domains, each of which relies on the communication of information deemed important. Although blinded to the scores of others, raters were not blinded to the LLM-generated source of the summaries, which could introduce rater bias and artificially inflate or deflate scores. They were also not blinded to the study hypothesis, which could introduce confirmation bias as raters could subconsciously inflate scores to align with the goals of the study.
Although infrequent, there were a few substantial disagreements between raters on the quality of summaries, with slight weighted agreement. The reported mean scores for each of the domains must be interpreted with caution given discordant ratings. It is also important to note that the observed slight agreement may partially reflect a statistical ceiling effect as 74.8% (371/496) of the total ratings were 5/5, and κ can underestimate agreement when ratings cluster within a narrow range and chance agreement is inflated. It is unclear whether the remaining disagreement represents inadequate training of raters, unclear definitions, or a limitation of using the PDQI for this task. That one particular rater often provided lower ratings than their counterparts may reflect lenience on the part of the other raters, particularly in light of the possibility of confirmation bias resulting from the fact that raters were unblinded to the study hypothesis. Prior to deployment in clinical practice, in addition to larger transfer call sample sizes and proportionally larger numbers of summary raters, follow-up studies could include additional raters per summary to more robustly characterize the frequency and magnitude of outliers in the domain of interrater disparities.
The original 9-item PDQI (PDQI-9) was developed and validated for the evaluation of notes across 9 criteria. At the conception of this study, we identified an unvalidated adaptation of the PDQI-9 that assessed the quality of LLM-generated text. Since then, there has been a validated evaluation of similar adaptations of the PDQI-9 for LLM-generated text that include the addition of the “hallucination free” domain and the removal of the “up-to-date” domain [19]. There has also been validation of a similar adapted PDQI-9 [20] for LLM-generated summaries that removed the “up-to-date” and “consistent” domains and included the assessment of hallucination in the “accuracy” domain. Notably, the latter defined hallucinations as falsifications (information or data distorted from the original note) or fabrications (made-up information or data that could be plausible but based on nonexistent facts) present in the summary. Regardless, the use of an unvalidated altered instrument may undermine the reliability of the scores to measure true summary quality.
To make this type of automated call summary tool usable in day-to-day clinical practice, the individual components (ie, call transcription and LLM analysis) need to be integrated such that summary generation is transparent to the summary recipient. In addition, the summary would need to be uploaded to the patient’s chart seamlessly, ideally without any incremental effort from the accepting or receiving clinician. Although implementation would require IT support, the use of ambient AI to document outpatient clinic visits [21,22] suggests that this objective is readily achievable in the tertiary medical center environment.