Identifying Deprescribing Opportunities With Large Language Models in Older Adults: Retrospective Cohort Study
Open accessPublished version of a preprint
Abstract
Background: Polypharmacy, the concurrent use of multiple medications, is prevalent among older adults and associated with increased risks for adverse drug events including falls. Deprescribing, the systematic process of discontinuing potentially inappropriate medications, aims to mitigate these risks. However, the practical application of deprescribing criteria in emergency settings remains limited due to time constraints and criteria complexity. Objective: This study aims to evaluate the performance of a large language model (LLM)–based pipeline in identifying deprescribing opportunities for older emergency department (ED) patients with polypharmacy, using 3 different sets of criteria: Beers, Screening Tool of Older People’s Prescriptions, and Geriatric Emergency Medication Safety Recommendations. The study further evaluates LLM confidence calibration and its ability to improve recommendation performance. Methods: We conducted a retrospective cohort study of older adults presenting to an ED in a large academic medical center in the Northeast United States from January 2022 to March 2022. A random sample of 100 patients (712 total oral medications) was selected for detailed analysis. The LLM pipeline consisted of two steps: (1) filtering high-yield deprescribing criteria based on patients’ medication lists, and (2) applying these criteria using both structured and unstructured patient data to recommend deprescribing. Model performance was assessed by comparing model recommendations to those of trained medical students, with discrepancies adjudicated by board-certified ED physicians. Selective prediction, a method that allows a model to abstain from low-confidence predictions to improve overall reliability, was applied to assess the model’s confidence and decision-making thresholds. Results: The LLM was significantly more effective in identifying deprescribing criteria (positive predictive value: 0.83; negative predictive value: 0.93; McNemar test for paired proportions: χ21=5.985; P=.02) relative to medical students, but showed limitations in making specific deprescribing recommendations (positive predictive value=0.47; negative predictive value=0.93). Adjudication revealed that while the model excelled at identifying when there was a deprescribing criterion related to one of the patient’s medications, it often struggled with determining whether that criterion applied to the specific case due to complex inclusion and exclusion criteria (54.5% of errors) and ambiguous clinical contexts (eg, missing information; 39.3% of errors). Selective prediction only marginally improved LLM performance due to poorly calibrated confidence estimates. Conclusions: This study highlights the potential of LLMs to support deprescribing decisions in the ED by effectively filtering relevant criteria. However, challenges remain in applying these criteria to complex clinical scenarios, as the LLM demonstrated poor performance on more intricate decision-making tasks, with its reported confidence often failing to align with its actual success in these cases. The findings underscore the need for clearer deprescribing guidelines, improved LLM calibration for real-world use, and better integration of human–artificial intelligence workflows to balance artificial intelligence recommendations with clinician judgment.
AI snapshot
Written from the paper’s full text.
TL;DR
In a retrospective cohort of older emergency department patients with polypharmacy, a GPT-4o pipeline was better than medical students at filtering medications for relevant deprescribing criteria but weaker at making specific deprescribing recommendations. Selective prediction only marginally improved performance because the model’s confidence was poorly calibrated.
Key findings
- For criteria filtering, GPT-4o achieved positive predictive value 0.83, negative predictive value 0.93, and F1-score 0.86 (precision 0.83, recall 0.90), significantly outperforming medical students (80.1% vs 59.5% correct; McNemar χ²₁=5.985; P=.02).
- For specific deprescribing recommendations, performance was lower: positive predictive value 0.47, negative predictive value 0.93, and F1-score 0.58 (precision 0.47, recall 0.76); the difference from medical students was not statistically significant (P=.178).
- Adjudication of 126 discrepancies found that 54.5% of errors involved complex inclusion/exclusion criteria and 39.3% involved ambiguous clinical contexts such as missing information. The LLM recommended deprescribing more often than medical students (11.6% vs 1.91%).
- Selective prediction only marginally improved performance; confidence was poorly calibrated, with minimum confidence 58.5% for eligibility filtering and 54.5% for deprescribing recommendations.
Methods
- Retrospective cohort study of adults aged ≥65 years with polypharmacy (≥5 active outpatient medications) presenting to an emergency department in a large academic medical center in the Northeast United States from January 2022 to March 2022. A random sample of 100 patients (712 oral medications) was selected; after prompt development, the final evaluation cohort was 92 patients and 626 medications.
- The GPT-4o pipeline used two stages: first filtering high-yield deprescribing criteria based on the patient’s medication list, then applying those criteria using structured and unstructured EHR data to recommend deprescribing. Recommendations were compared with those of trained medical students, with discrepancies adjudicated by board-certified emergency department physicians.
- Analyses included McNemar tests for paired proportions, error classification, and selective prediction using verbalized confidence and consistency-based confidence, evaluated with risk-coverage curves substituting F1-score for risk.
Limitations
- The paper states that the retrospective EHR design may not capture real-time emergency decision-making, the single academic medical center limits generalizability, selective prediction was not universally effective, and reliance on textual prompts and structured EHR data may miss nuanced clinical context.
- The paper states that the small sample size (100 patients) limits statistical power, API costs were approximately US $300-$400 for 100 patients and would scale linearly, medical student annotations may introduce variability despite physician adjudication, no geriatricians were explicitly included in criteria selection, and the selected criteria may not encompass all deprescribing scenarios.
- Evident from the text: the final evaluation cohort was 92 patients rather than the initially sampled 100, and the study lacked a prior power calculation because of limited prior evidence on LLM performance in this task.
Model-written summary. Check it against the paper before relying on it.
Integrity signals
Registry
Cites
Statistics
Claims the paper makes
0 of 26 cited sources checked · 0 supported · 0 flagged
- Findingown data
In the second step-making specific deprescribing recommendations-the LLM encountered considerable difficulty, particularly when dealing with ambiguous criteria, missing information, and nuanced clinical scenarios.
“in the second step—making specific deprescribing recommendations—the LLM encountered considerable difficulty, particularly when dealing with ambiguous criteria, missing information, and nuanced clinical scenarios.”
Quote in the paperSubstantial claim - Findingown data
The model excelled at identifying when there was a deprescribing criterion related to one of the patient's medications.
“Adjudication revealed that while the model excelled at identifying when there was a deprescribing criterion related to one of the patient’s medications”
Quote in the paperSubstantial claim - Findingown data
A major source of discrepancy was the significantly higher likelihood of the large language model to recommend deprescribing (11.6%) compared to the medical students (1.91%).
“Notably, the confusion matrix (Figure 4) revealed that a major source of discrepancy was the significantly higher likelihood of the LLM to recommend deprescribing (11.6%) compared to the medical students (1.91%).”
Quote in the paperSubstantial claim - Findingown data
Feasibility results by criteria list showed more likelihood for deprescribing in both outpatient and inpatient contexts compared to the emergency department across all 3 criteria lists.
“We also present the results of feasibility in various clinical settings by criteria list in Figure S1 in Multimedia Appendix 1, showing more likelihood for deprescribing in both outpatient and inpatient contexts, compared to the ED, across all 3 criteria lists.”
Quote in the paperNot a substantial claim - Findingown data
The model often struggled with determining whether a deprescribing criterion applied to the specific case due to ambiguous clinical contexts (eg, missing information; 39.3% of errors).
“and ambiguous clinical contexts (eg, missing information; 39.3% of errors).”
Quote in the paperSubstantial claim
- Findingown data
The LLM outperformed medical students by a significant margin (80.1% vs 59.5% correct, McNemar test: P=.02).
“the LLM outperformed medical students by a significant margin (80.1% vs 59.5% correct, McNemar test: P=.02).”
Quote in the paperSubstantial claim- Rests on an assumption: Medical students’ performance is a valid benchmark for evaluating the LLM’s clinical reasoning in this task.
- Findingown data
The LLM is poorly calibrated.
“However, despite some minor improvements, we find that the LLM is poorly calibrated, as shown in Figure 6B.”
Quote in the paperSubstantial claim - Findingown data
The large language model was significantly more effective in identifying deprescribing criteria (positive predictive value: 0.83; negative predictive value: 0.93; McNemar test: χ21=5.985; P=.02) relative to medical students.
“The LLM was significantly more effective in identifying deprescribing criteria (positive predictive value: 0.83; negative predictive value: 0.93; McNemar test for paired proportions: χ21=5.985; P=.02) relative to medical students”
Quote in the paperSubstantial claim - Findingown data
The large language model effectively identified cases lacking relevant high-yield deprescribing criteria, achieving an F1-score of 0.86 (precision=0.83, recall=0.90).
“The LLM effectively identified these cases, achieving an F1-score of 0.86 (precision=0.83, recall=0.90).”
Quote in the paperSubstantial claim - Findingown data
The large language model's reported confidence often fails to align with its actual success in intricate decision-making tasks.
“with its reported confidence often failing to align with its actual success in these cases.”
Quote in the paperSubstantial claim - Findingreasoning
The model often recommended deprescribing medications that, while potentially inappropriate, required specific contextual qualifiers to justify deprescribing-criteria that the LLM misapplied due to lack of explicit context or ambiguous language in the guidelines.
“Additionally, the model often recommended deprescribing medications that, while potentially inappropriate, required specific contextual qualifiers (eg, patient’s life expectancy, nutritional status, frailty status) to justify deprescribing—criteria that the LLM misapplied due to lack of explicit context or ambiguous language in the guidelines.”
Quote in the paperSubstantial claim - Findingown data
Across all discrepancies, GPT-4o was significantly more effective in criteria filtering compared to the medical student (McNemar test: χ21=5.985; P=.015).
“Across all discrepancies, GPT-4o was significantly more effective in criteria filtering compared to the medical student (McNemar test for paired proportions [30]: χ21=5.985; P=.015).”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [30] source
Not checked in the free run.
- Findingown data
The LLM frequently overrecommended deprescribing compared to medical students.
“The LLM’s frequent tendency to overrecommend deprescribing, as compared to medical students, indicates that clear boundaries for medication eligibility and exclusion are critical for reducing false positives in automated recommendations.”
Quote in the paperSubstantial claim- Rests on an assumption: Medical students’ performance is a valid benchmark for evaluating the LLM’s clinical reasoning in this task.
- Findingown data
Adjudication was chosen as the gold standard for resolving discrepancies because preliminary research showed low interrater reliability among medical students for both steps of the process (Cohen k: step 1=0.68, step 2=0.33) across 75 medications.
“Adjudication was chosen as the gold standard for resolving discrepancies because preliminary research showed low IRR among medical students for both steps of the process (Cohen k: step 1=0.68, step 2=0.33) across 75 medications.”
Quote in the paperSubstantial claim - Findingown data
The LLM displayed a minimum confidence level of 54%, even in cases where its recommendations were incorrect.
“Specifically, the LLM displayed a minimum confidence level of 54%, even in cases where its recommendations were incorrect.”
Quote in the paperSubstantial claim - Findingown data
315 medications (50.3% of the total) lacked relevant high-yield deprescribing criteria.
“As shown in Figure 4, 315 medications (50.3% of the total) lacked relevant high-yield deprescribing criteria.”
Quote in the paperSubstantial claim - Findingown data
The LLM demonstrated strengths in the initial filtering step, accurately identifying a high proportion of medications that matched deprescribing criteria.
“The LLM demonstrated strengths in the initial filtering step, accurately identifying a high proportion of medications that matched deprescribing criteria, thus offering the potential to support clinicians in rapidly screening complex medication lists.”
Quote in the paperSubstantial claim - Findingown data
Cases with the potential for human harm were observed, such as suggesting deprescribing anticoagulation in a patient with recent thromboembolism.
“Cases with the potential for human harm were observed, such as suggesting deprescribing anticoagulation in a patient with recent thromboembolism.”
Quote in the paperSubstantial claim - Findingown data
The large language model showed limitations in making specific deprescribing recommendations (positive predictive value=0.47; negative predictive value=0.93).
“but showed limitations in making specific deprescribing recommendations (positive predictive value=0.47; negative predictive value=0.93).”
Quote in the paperSubstantial claim - Findingown data
Selective prediction only marginally improved large language model performance due to poorly calibrated confidence estimates.
“Selective prediction only marginally improved LLM performance due to poorly calibrated confidence estimates.”
Quote in the paperNot a substantial claim - Findingreasoning
Selective prediction marginally improved the LLM's filtering accuracy by enabling it to defer uncertain cases to human reviewers.
“Selective prediction marginally improved the LLM’s filtering accuracy by enabling it to defer uncertain cases to human reviewers.”
Quote in the paperSubstantial claim - Findingnone given
Cases were seen in which the LLM recommended deprescribing without citing a specific criterion.
“Additionally, cases were seen in which the LLM recommended deprescribing without citing a specific criterion.”
Quote in the paperSubstantial claim - Findingown data
Consistency-based selective prediction demonstrates a positive linear relationship between accuracy in deprescribing recommendations and deferring fractions.
“consistency-based selective prediction demonstrates a positive linear relationship between accuracy in deprescribing recommendations and deferring fractions.”
Quote in the paperNot a substantial claim - Findingown data
In the second step, applying relevant criteria to make a recommendation, the large language model performed less effectively, with an F1-score of 0.58 (precision=0.47, recall=0.76).
“In this second step, which involved applying relevant criteria to make a recommendation, the LLM performed less effectively, with an F1-score of 0.58 (precision=0.47, recall=0.76).”
Quote in the paperSubstantial claim - Findingown data
Reduced inclusion of STOPP criteria was primarily attributed to panelists' concerns that the information required was not readily accessible within the electronic health record and would necessitate additional data at the point of care.
“Reduced inclusion of STOPP criteria was primarily attributed to panelists’ concerns that the information required was not readily accessible within the EHR and would necessitate additional data at the point of care.”
Quote in the paperNot a substantial claim - Findingown data
Among medications with relevant criteria, 64 (10.2% of the total) were cases where either GPT-4o or the medical student recommended deprescribing.
“Among those medications with relevant criteria, 64 (10.2% of the total) were cases where either GPT-4o or the medical student recommended deprescribing.”
Quote in the paperSubstantial claim - Findingown data
The model often struggled with determining whether a deprescribing criterion applied to the specific case due to complex inclusion and exclusion criteria (54.5% of errors).
“it often struggled with determining whether that criterion applied to the specific case due to complex inclusion and exclusion criteria (54.5% of errors)”
Quote in the paperSubstantial claim - Premisecitation
Polypharmacy can increase the risk of negative consequences for patients, including emergency department visits, adverse drug events, falls, disability, and inappropriate medication use.
“Although necessary and beneficial for some patients, polypharmacy can increase the risk of negative consequences for patients, including ED visits, adverse drug events, falls, disability, and inappropriate medication use [1].”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [1] Halli-Tierney, AD, 2019, Am Fam Physician
Not checked in the free run.
- Premisecitation
Explicit criterion-based deprescribing assessments such as STOPP and Beers are often impractical to implement in time-constrained clinical settings such as the emergency department due to the need to evaluate multiple clinical indications and specialist-prescribed medications.
“These explicit assessments are criterion-based with clear standards but are often impractical to implement in time-constrained clinical settings, such as ED, due to the need to evaluate multiple clinical indications and specialist-prescribed medications [10].”
- Also stated in: § Background · plain
“However, the practical application of deprescribing criteria in emergency settings remains limited due to time constraints and criteria complexity.”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [10] source
Not checked in the free run.
- Also stated in: § Background · plain
- Premisecitation
The majority of clinical reasoning evaluations on large language models have been conducted using standardized exams or digital case reports.
“However, the majority of clinical reasoning evaluations on LLMs have been conducted using standardized exams (the United States Medical Licensing Examination) or digital case reports [14,19].”
Quote in the paperSubstantial claim0 of 2 sources checked - Premisecitation
The need to filter deprescribing criteria before proceeding was identified in preliminary research, which revealed that one of the main causes of discrepancies between physicians and large language models arose from ambiguous inclusion or exclusion conditions in deprescribing criteria.
“The need to filter the criteria before proceeding with the study was identified in our preliminary research [20], which revealed that one of the main causes of discrepancies between physicians and LLMs arose from ambiguous inclusion or exclusion conditions in deprescribing criteria.”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [20] source
Not checked in the free run.
- Premisecitation
Large language models have been shown to extract medication-related data such as medication name, dosage, and frequency, necessary for the application of deprescribing criteria.
“Moreover, they have been shown to extract medication-related data such as medication name, dosage, and frequency, necessary for the application of deprescribing criteria [17].”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [17] Goel, A, 2023, Proc Mach Learn Res
Not checked in the free run.
- Premisecitation
Attempts to digitize deprescribing criteria into electronic clinical decision support have raised difficulties, typically requiring a labor-intensive coding process and unstructured information such as free text from patient records to contextualize certain criteria.
“Attempts to digitize these criteria into electronic clinical decision support (CDS) have raised difficulties, typically requiring a labor-intensive coding process and unstructured information such as free text from patient records to contextualize certain criteria [11,12].”
Quote in the paperSubstantial claim0 of 2 sources checked - Interpretationown data
Our results contradict recent work that suggests that consistency-based methods alleviate some concerns about overconfidence.
“While verbalized confidence is known to be overconfident in clinical question answering [38], our results contradict recent work that suggests that consistency-based methods alleviate some of these concerns [39].”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [39] source
Not checked in the free run.
- Study design and conductcitation
GPT-4o's decision confidence was collected using 2 validated confidence elicitation methods: chain-of-thought verbalized confidence and self-random sampling with average-confidence aggregation, referred to as consistency-based confidence.
“To do so, we collected GPT-4o’s decision confidence for both steps using 2 validated confidence elicitation methods: chain-of-thought verbalized confidence and self-random sampling with average-confidence aggregation, referred to as consistency-based confidence [24].”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [24] XiongMHuZLuXLiYFuJHeJHooiBCan LLMs Express Their Uncertainty? An Empirical Evaluation of Confidence Elicitation in LLMsCan LLMs express their uncertainty? An empirical evaluation of confidence elicitation in LLMs2024The Twelfth International Conference on Learning RepresentationsMay 7, 2024Vienna, Austria
Not checked in the free run.
- Interpretationreasoning
The model's performance in deprescribing recommendations highlights the challenge of translating structured criteria into actionable clinical decisions, especially when faced with ambiguous inclusion or exclusion conditions.
“The model’s performance in these recommendations highlights the challenge of translating structured criteria into actionable clinical decisions, especially when faced with ambiguous inclusion or exclusion conditions.”
Quote in the paperSubstantial claim - Interpretationown data
The LLM has a tendency toward overconfidence, particularly in its deprescribing recommendations.
“This indicates a tendency toward overconfidence, particularly in its deprescribing recommendations.”
Quote in the paperSubstantial claim - Study design and conductown data
The gold standard for evaluating whether the large language model or the medical student was correct was derived from senior annotator (board-certified emergency department physicians) adjudication of discrepancies.
“We evaluated whether the LLM or the medical student was correct, using a gold standard derived from senior annotator (board-certified ED physicians) adjudication of discrepancies.”
Quote in the paperSubstantial claim- Rests on an assumption: Board-certified emergency department physician adjudication of discrepancies is a valid gold standard for determining correct deprescribing recommendations.
- Interpretationown data
Large language models can support deprescribing decisions in the emergency department by effectively filtering relevant criteria.
“This study highlights the potential of LLMs to support deprescribing decisions in the ED by effectively filtering relevant criteria.”
Quote in the paperSubstantial claim - Interpretationreasoning
This work highlights the limitations of current LLMs in complex, context-sensitive clinical decision-making tasks.
“However, this work also highlights the limitations of current LLMs in complex, context-sensitive clinical decision-making tasks.”
Quote in the paperSubstantial claim - Interpretationown data
Challenges remain in applying deprescribing criteria to complex clinical scenarios, as the large language model demonstrated poor performance on more intricate decision-making tasks.
“However, challenges remain in applying these criteria to complex clinical scenarios, as the LLM demonstrated poor performance on more intricate decision-making tasks”
Quote in the paperSubstantial claim - Study design and conductreasoning
The selective prediction methods were not universally effective, particularly in the nuanced task of deprescribing recommendations.
“Second, the selective prediction methods, while providing insights into the LLM’s confidence, were not universally effective, particularly in the nuanced task of deprescribing recommendations.”
Quote in the paperNot a substantial claim - Study design and conductnone given
The study's focus on a single large academic medical center limits the generalizability of the findings to other settings with different patient populations, documentation patterns, and health care practices.
“The study’s focus on a single large academic medical center limits the generalizability of our findings to other settings with different patient populations, documentation patterns, and health care practices.”
Quote in the paperNot a substantial claim - Interpretationreasoning
Despite successes in eligibility filtering, challenges remain when applying complex criteria to specific patient cases.
“However, despite successes in eligibility filtering, challenges remain when applying complex criteria to specific patient cases.”
Quote in the paperSubstantial claim - Interpretationown data
The effectiveness of selective prediction was limited by the poorly calibrated confidence levels assigned by the LLM to its decisions.
“However, the effectiveness of this approach was limited by the poorly calibrated confidence levels assigned by the LLM to its decisions.”
Quote in the paperSubstantial claim - Premisecitation
One study used a similar "filter-and-apply" pipeline, in which trials were first filtered and then matched to patients, showcasing the effectiveness of this approach.
“One study used a similar “filter-and-apply” pipeline, in which trials were first filtered and then matched to patients, showcasing the effectiveness of this approach [35].”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [35] source
Not checked in the free run.
- Interpretationown data
The adjudication, combined with strong overall performance (maximum F1-score: 87.8%) using selective prediction methods, suggests that the LLM can effectively minimize the number of criteria requiring final review for deprescribing recommendations.
“The adjudication, combined with strong overall performance (maximum F1-score: 87.8%) using selective prediction methods, suggests that the LLM can effectively minimize the number of criteria requiring final review for deprescribing recommendations.”
Quote in the paperSubstantial claim- Rests on an assumption: Senior clinician adjudication provides a valid gold standard for resolving discrepancies between GPT-4o and medical students.
- Interpretationown data
This study demonstrates the potential of LLMs to augment clinical decision support by effectively filtering deprescribing criteria for older adults with polypharmacy in ED.
“This study demonstrates the potential of LLMs to augment clinical decision support by effectively filtering deprescribing criteria for older adults with polypharmacy in ED.”
Quote in the paperSubstantial claim
- May combine two claims. 24
- Classified as probably not a substantial claim. 22
May combine two claims.
Findingown dataThe confidence distribution is severely left-skewed with the minimum confidence being 58.5% in eligibility filtering and 54.5% in deprescribing recommendations.
“Despite consistency-based weighting, the confidence distribution is severely left-skewed with the minimum confidence being 58.5% in eligibility filtering and 54.5% in deprescribing recommendations.”
Quote in the paperNot a substantial claimMay combine two claims.
Findingown dataAfter filtering out nonoral medications, there were 712 total oral medications across the cohort, with a median of 6 oral medications per patient.
“As our criteria only pertain to oral medication, nonoral medications were subsequently filtered out, resulting in 712 total oral medications across the cohort and a median of 6 oral medications per patient (Figure 2).”
Quote in the paperNot a substantial claimMay combine two claims.
Findingown dataConsistency-based confidence exhibited broader and higher F1-score ranges, with step 1 spanning 0.85 to 0.88 and step 2 ranging from 0.62 to 0.73.
“consistency-based confidence exhibited broader and higher F1-score ranges, with step 1 spanning 0.85 to 0.88 and step 2 ranging from 0.62 to 0.73 (Figure 6A).”
Quote in the paperSubstantial claimMay combine two claims.
Findingown dataVerbalized confidence demonstrated a narrower range of F1-scores overall, with step 1 (eligibility) ranging from 0.860 to 0.863 and step 2 (deprescribing) from 0.58 to 0.69.
“Verbalized confidence demonstrated a narrower range of F1-scores overall, with step 1 (eligibility) ranging from 0.860 to 0.863 and step 2 (deprescribing) from 0.58 to 0.69, as shown in Figure S3 in Multimedia Appendix 1.”
Quote in the paperSubstantial claimMay combine two claims.
Findingown dataIn applying relevant criteria, GPT-4o performed worse than the medical students, though the difference was not statistically significant (McNemar test: χ21=1.818; P=.178).
“However, in applying relevant criteria, GPT-4o performed worse than the medical students, though the difference was not statistically significant (McNemar test: χ21=1.818; P=.178).”
Quote in the paperNot a substantial claimClassified as probably not a substantial claim. May combine two claims.
FindingreasoningFrom the initial 100 patients, 10 were set aside for prompt engineering and calculation of interrater reliability between junior annotators; because fewer iterations were needed, the remaining 2 patients were included in the final study cohort, resulting in a final evaluation cohort of 92 patients and 626 medications.
“From our initial study cohort of 100 patients, 10 patients were set aside for both prompt engineering and calculation of IRR between junior annotators. Fewer iterations were needed to refine the prompt than initially anticipated, so the remaining 2 patients were included in the final study cohort. This resulted in a final evaluation cohort of 92 patients, encompassing a total of 626 medications.”
Quote in the paperSubstantial claimMay combine two claims.
FindingreasoningAmbiguities in criteria definitions, such as those related to medication administration routes and drug classes, present substantial barriers to automation and contribute to discrepancies between human and model interpretations.
“Ambiguities in criteria definitions, such as those related to medication administration routes and drug classes, present substantial barriers to automation and contribute to discrepancies between human and model interpretations.”
Quote in the paperSubstantial claimMay combine two claims.
Findingown dataOn average, STOPP criteria had the lowest clinical risk and electronic health record computability ratings, while Beers criteria had the highest, contributing to their respective adoption rates of 45.9% and 62.5% among high-yield criteria.
“On average, STOPP criteria had the lowest clinical risk and EHR computability ratings, while the Beers criteria had the highest, contributing to their respective adoption rates of 45.9% and 62.5% among high-yield criteria.”
Quote in the paperNot a substantial claimMay combine two claims.
Findingown dataIn total, 10,977 unique patients across 15,161 emergency department encounters from January 2022 to March 2022 met the selection criteria, from which 100 patients were randomly selected.
“In total, we identified 10,977 unique patients across 15,161 emergency department encounters from January 2022 to March 2022 meeting our selection criteria, from which 100 patients were randomly selected (Table 2).”
Quote in the paperNot a substantial claimMay combine two claims.
Findingown dataGPT-4o recommended deprescribing without access to current electrolyte values, instead basing its suggestion on a history of chronic kidney disease, a condition associated with potential electrolyte imbalances but not meeting the relevant inclusion criteria.
“GPT-4o recommended deprescribing without access to current electrolyte values, instead basing its suggestion on a history of chronic kidney disease, a condition associated with potential electrolyte imbalances but not meeting the relevant inclusion criteria.”
Quote in the paperSubstantial claimClassified as probably not a substantial claim.
Findingown dataFor cases where the large language model and medical students differed, 2 senior annotators adjudicated 126 discrepancies after standardizing the codebook and verifying interrater reliability (Cohen k: eligibility=0.795, deprescribing=0.745).
“For cases where the LLM and the medical students differed, 2 senior annotators (board-certified Emergency Medicine physicians) adjudicated 126 discrepancies after standardizing the codebook and verifying IRR (Cohen k: eligibility=0.795, deprescribing=0.745).”
Quote in the paperSubstantial claimMay combine two claims.
Findingown dataCases of misclassification were relatively uncommon and primarily related to nonstandard drug class names or overly broad groups.
“Cases of misclassification were relatively uncommon and primarily related to nonstandard drug class names or overly broad groups, which could be improved with refined deprescribing criteria.”
Quote in the paperNot a substantial claimMay combine two claims.
PremisecitationLarge language models are excellent in-context learners, requiring very little labeled data to make predictions.
“Finally, LLMs are excellent in-context learners, requiring very little labeled data to make predictions [18].”
Quote in the paperSubstantial claim0 of 1 sources checked- cites [18] Agrawal, M, 2022, Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing source
Not checked in the free run.
Classified as probably not a substantial claim.
PremisereasoningThe ability of large language models to perform clinical reasoning and calibrate responses over physician-generated text remains understudied.
“Their ability to perform clinical reasoning and calibrate responses over physician-generated text remains understudied.”
Quote in the paperSubstantial claimClassified as probably not a substantial claim.
Study design and conductreasoningDue to budgetary constraints and a lack of prior evidence regarding large language model performance in this task to guide a power calculation, a random sample of 100 unique patients was selected for evaluation.
“Due to budgetary constraints and a lack of prior evidence regarding the performance of LLMs in this task to guide a power calculation, we selected a random sample of 100 unique patients for evaluation.”
Quote in the paperSubstantial claimClassified as probably not a substantial claim.
PremisecitationLarge language models have been shown to interpret complex clinical situations and offer recommendations, from differential diagnoses to care management.
“Large language models (LLMs) have been shown to interpret complex clinical situations and offer recommendations, from differential diagnoses to care management, leading to growing interest in their application in the medical field [13-16].”
Quote in the paperSubstantial claim0 of 4 sources checkedProbably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.
Study design and conductnone givenThe process for selecting the criteria did not explicitly include any geriatricians, though it did include a range of individuals who regularly care for older adults.
“Our process for selecting these criteria also did not explicitly include any geriatricians, though did include a range of individuals who regularly care for older adults.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing).
Study design and conductnone givenThe model's reliance on textual prompts and structured EHR data may not fully account for nuanced clinical contexts that influence deprescribing decisions.
“Additionally, the model’s reliance on textual prompts and structured EHR data may not fully account for nuanced clinical contexts that influence deprescribing decisions.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Study design and conductnone givenAll large language model calls were performed with a fixed temperature of 0 and a seed to ensure reproducibility and deterministic outputs.
“Aside from the consistency-based method described later, all LLM calls were performed with a fixed temperature (temperature=0; low randomness in generated responses) and seed to ensure reproducibility and deterministic outputs.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
PremisecitationSelective prediction has been commonly found to improve performance in non-text-based applications.
“This human-in-the-loop decision-making pipeline is known as selective prediction and has been commonly found to improve performance in non–text-based applications [25,26].”
Quote in the paperNot a substantial claim0 of 2 sources checkedProbably not a substantial claim: may not pass S3 (load-bearing), S4 (based).
Interpretationnone givenThe lack of standardized guidelines for implementing deprescribing criteria in LLMs poses a challenge to consistency and accuracy.
“The lack of standardized guidelines for implementing deprescribing criteria in LLMs also poses a challenge to consistency and accuracy.”
Quote in the paperNot a substantial claimMay combine two claims.
PremisecitationSeparating criteria filtering and application reduces confusion due to large input context sizes and ensures that extraneous context does not distract the large language model.
“Separating the process of criteria filtering and application both reduces confusion due to large input context sizes [22] and ensures that extraneous context does not distract the LLM [23].”
Quote in the paperSubstantial claim0 of 2 sources checked- cites [22] source
Not checked in the free run.
- cites [23] ShiFChenXMisraKScalesNDohanDChiESchärliNZhouDLarge language models can be easily distracted by irrelevant context2023ICML'23: Proceedings of the 40th International Conference on Machine LearningJuly 23, 2023Honolulu, HI
Not checked in the free run.
Probably not a substantial claim: may not pass S3 (load-bearing).
Study design and conductcitationErrors leading to incorrect recommendations by the large language model were classified using a prior evaluation framework into 4 error types: incorrect reading comprehension, incorrect recall of knowledge, incorrect reasoning step, and not enough information.
“Finally, we classified the errors leading to incorrect recommendations by the LLM, leveraging a prior evaluation framework [29]. We classified each error as 1 of 4 error types: incorrect reading comprehension, incorrect recall of knowledge, incorrect reasoning step, and not enough information, as described in Table 1.”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [29] source
Not checked in the free run.
Probably not a substantial claim: may not pass S3 (load-bearing).
PremisecitationVerbalized confidence is known to be overconfident in clinical question answering.
“While verbalized confidence is known to be overconfident in clinical question answering [38], our results contradict recent work that suggests that consistency-based methods alleviate some of these concerns [39].”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [38] source
Not checked in the free run.
Probably not a substantial claim: may not pass S3 (load-bearing).
Study design and conductnone givenThe retrospective nature of the analysis, relying on historical data from EHRs, may not fully capture the complexity of real-time clinical decision-making in emergency settings.
“First, the retrospective nature of our analysis, relying on historical data from EHRs, may not fully capture the complexity of real-time clinical decision-making in emergency settings.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S2 (checkable), S3 (load-bearing). May combine two claims.
InterpretationreasoningIf implemented in clinical decision support, these inaccuracies might contribute to increased alert fatigue and extend the time required to interpret LLM-generated recommendations, potentially offsetting the intended efficiency gains in identifying deprescribing opportunities.
“If implemented in clinical decision support, these inaccuracies might contribute to increased alert fatigue and extend the time required to interpret LLM-generated recommendations, potentially offsetting the intended efficiency gains in identifying deprescribing opportunities.”
Quote in the paperNot a substantial claimMay combine two claims.
InterpretationcitationOur findings on the LLM's performance in identifying medications with relevant deprescribing criteria based on eligibility guidelines align with evidence from clinical trial matching literature, where LLMs have shown performance comparable to physicians in applying such criteria to identify eligible patients.
“Our findings on the LLM’s performance in identifying medications with relevant deprescribing criteria based on eligibility guidelines align with evidence from clinical trial matching literature [34], where LLMs have shown performance comparable to physicians in applying such criteria to identify eligible patients.”
Quote in the paperNot a substantial claim0 of 1 sources checked- cites [34] source
Not checked in the free run.
Probably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.
Study design and conductnone givenThe criteria used (STOPP, Beers, and GEMS-Rx) were selected based on their perceived clinical risk and EHR computability, which may not encompass all relevant deprescribing scenarios.
“Finally, the criteria used (STOPP, Beers, and GEMS-Rx) were selected based on their perceived clinical risk and EHR computability, which may not encompass all relevant deprescribing scenarios.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Study design and conductown dataInterrater reliability was measured between the 2 senior emergency department physicians prior to adjudication on the full set of discrepancies.
“Similarly to the junior annotators, we measured the IRR between the 2 senior ED physicians, prior to adjudication on the full set of discrepancies.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing).
Study design and conductreasoningEach medication was evaluated individually to prevent errors from simultaneous processing, such as misattribution of criteria or medication omissions.
“Each medication was evaluated individually to prevent errors from simultaneous processing, such as misattribution of criteria or medication omissions.”
Quote in the paperNot a substantial claimProbably not a substantial claim: may not pass S3 (load-bearing). May combine two claims.
Study design and conductreasoningThe study relied on medical students for initial annotation; while annotations were reviewed and adjudicated by board-certified physicians, this process may introduce variability, potentially affecting the reliability of their use as a gold standard in selective prediction methods.
“Additionally, the study relied on medical students for initial annotation. While these annotations were reviewed and adjudicated by board-certified physicians, this process may introduce variability, potentially affecting the reliability of their use as a gold standard in selective prediction methods.”
Quote in the paperNot a substantial claimMay combine two claims.
Study design and conductnone givenThe small sample size for detailed analysis (100 patients) limits the statistical power and may not reflect broader patterns of medication use and deprescribing needs.
“Third, the small sample size for detailed analysis (100 patients) limits the statistical power and may not reflect broader patterns of medication use and deprescribing needs.”
Quote in the paperNot a substantial claimMay combine two claims.
Interpretationown dataConsistency-based confidence provides more flexibility and improved performance compared to verbalized confidence across both steps.
“These results suggest that consistency-based confidence provides more flexibility and improved performance compared to verbalized confidence across both steps.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S2 (checkable).
InterpretationreasoningA notable finding from this study is the need for clearer and more consistent deprescribing guidelines.
“A notable finding from this study is the need for clearer and more consistent deprescribing guidelines.”
Quote in the paperNot a substantial claimClassified as probably not a substantial claim. May combine two claims.
Study design and conductown dataTwo trained senior medical students classified all medications in the test cohort using a 2-stage pipeline, after first computing interrater reliability on an adjudication set of 75 medications from 5 patients.
“Two trained senior medical students (M4) classified all medications in the test cohort using a 2-stage pipeline, after first computing interrater reliability (IRR) on an adjudication set of 75 medications from 5 patients.”
Quote in the paperSubstantial claimClassified as probably not a substantial claim.
Study design and conductnone givenDiscrepancies between the medical students and the large language model across both stages were adjudicated by 2 board-certified emergency department physicians.
“Discrepancies between the students (junior annotators) and the LLM across both stages were adjudicated by 2 board-certified ED physicians (senior annotators).”
Quote in the paperSubstantial claimMay combine two claims.
PremisecitationSimilar errors to those observed in our work have been reported, such as incorrectly identifying patients who meet partial criteria, or assuming a patient with breast cancer does not have lung cancer simply because it is not explicitly mentioned.
“Similar errors to those observed in our work have been reported, such as incorrectly identifying patients who meet partial criteria, or assuming a patient with breast cancer does not have lung cancer simply because it is not explicitly mentioned [36,37].”
Quote in the paperNot a substantial claim0 of 2 sources checkedMay combine two claims.
InterpretationreasoningThe LLM showed promise in identifying medications eligible for deprescribing but faced challenges in making nuanced deprescribing recommendations, underscoring the need for human oversight in AI-driven processes.
“While the LLM showed promise in identifying medications eligible for deprescribing, it faced challenges in making nuanced deprescribing recommendations, underscoring the need for human oversight in AI-driven processes.”
Quote in the paperSubstantial claimProbably not a substantial claim: may not pass S2 (checkable). May combine two claims.
InterpretationreasoningThe findings underscore the potential of LLMs in AI-enabled automated CDS tools for deprescribing while emphasizing the need to refine deprescribing criteria and establish clearer guidelines to support the integration of AI into clinical practice.
“The findings underscore the potential of LLMs in AI-enabled automated CDS tools for deprescribing while emphasizing the need to refine deprescribing criteria and establish clearer guidelines to support the integration of AI into clinical practice.”
Quote in the paperNot a substantial claim- Rests on an assumption: Current deprescribing criteria were not originally designed for direct implementation in CDS systems, but rather as general recommendations for prescribing physicians.
The paper relies on this without support.
Board-certified emergency department physician adjudication of discrepancies is a valid gold standard for determining correct deprescribing recommendations.
“We evaluated whether the LLM or the medical student was correct, using a gold standard derived from senior annotator (board-certified ED physicians) adjudication of discrepancies.”
“Adjudication was chosen as the gold standard for resolving discrepancies because preliminary research showed low IRR among medical students for both steps of the process (Cohen k: step 1=0.68, step 2=0.33) across 75 medications.”
The paper relies on this without support.
The random sample of 100 patients, and the final 92-patient cohort, is representative of older emergency department patients with polypharmacy, so the large language model performance findings generalize to that population.
“Due to budgetary constraints and a lack of prior evidence regarding the performance of LLMs in this task to guide a power calculation, we selected a random sample of 100 unique patients for evaluation.”
“In total, we identified 10,977 unique patients across 15,161 emergency department encounters from January 2022 to March 2022 meeting our selection criteria, from which 100 patients were randomly selected (Table 2).”
The paper relies on this without support.
The electronic health record data used, including structured and unstructured information, are sufficiently complete and accurate for applying deprescribing criteria.
“In stage 2, GPT-4o was prompted to use its previously filtered criteria list, along with structured (eg, demographics, lab values, vitals, and past medical history) and unstructured (most recent progress note and discharge summary) information, to determine if the patient satisfied any deprescribing criteria and the medication should be recommended for deprescribing.”
The paper relies on this without support.
Current deprescribing criteria were not originally designed for direct implementation in CDS systems, but rather as general recommendations for prescribing physicians.
“It is important to note that current deprescribing criteria were not originally designed for direct implementation in CDS systems, but rather as general recommendations for prescribing physicians.”
The paper relies on this without support.
Medical students’ performance is a valid benchmark for evaluating the LLM’s clinical reasoning in this task.
“the LLM outperformed medical students by a significant margin (80.1% vs 59.5% correct, McNemar test: P=.02).”
The paper relies on this without support.
Senior clinician adjudication provides a valid gold standard for resolving discrepancies between GPT-4o and medical students.
“Adjudication by senior clinicians was used to resolve discrepancies, and selective prediction methods were tested to improve the model’s reliability.”
Claims other papers make about this one
Each is checked against this paper’s own text: does it say what the citing paper says it does?
GPT-4 achieved 100% image reproducibility in medication identification from pill images when re-queried 2 weeks later.
“Conversely, GPT-4 achieved 100% image reproducibility in medication identification from pill images when re-queried 2 weeks later.78”
From Large Language Models in Pharmacist-Led Direct Patient Care: A Rapid Review of Published Evaluations (2026) · cites this paper as [78]
The Methods (Deprescribing Recommendations by GPT-4o; Selective Prediction Methods, page unavailable) describe medication-list and patient-record inputs, plus repeated sampling for deprescribing decisions. A search for pill images, medication identification, 100% reproducibility, and re-querying after two weeks found no passage reporting the claimed result. (LLM, conf=0.78)
Experimental — accuracy not yet measured
Self-check
Percentages, sums and reported differences re-computed from the paper’s own numbers.
Each contradiction between two passages listed here was confirmed by an independent check.
Table labels 100 patients as evaluated, but final evaluation includes 92
MinorTable 2 describes all 100 sampled patients as included in the evaluation, whereas the Results section identifies 92 patients as the final evaluation cohort. The table does not distinguish the original sample from the patients actually included in the final evaluation.
“Table: Table 2 Demographic overview of the 100 patients included in the evaluation (N=100).”
“This resulted in a final evaluation cohort of 92 patients, encompassing a total of 626 medications.”
Conflicting score thresholds for high-yield criteria
MinorThe Methods section requires a score greater than 3 on each dimension. The Results section instead says criteria were excluded if they scored less than 3 on both dimensions, which would retain criteria scoring exactly 3 or scoring below 3 on only one dimension.
“We selected the top 50% of criteria with an average score greater than 3 on both dimensions, calculated across all experts, as high-yield criteria.”
“Any criteria scoring less than 3 on both clinical applicability and EHR computability were excluded resulting in 161 criteria.”
The citation lists haven’t been fetched yet
References 38 · Cited by 17
Research works printed but not linked: 5
- Halli-Tierney, AD, 2019, Am Fam Physician, 100, 32
- Goel, A, 2023, Proc Mach Learn Res, 82
- ShiFChenXMisraKScalesNDohanDChiESchärliNZhouDLarge language models can be easily distracted by irrelevant context2023ICML'23: Proceedings of the 40th International Conference on Machine LearningJuly 23, 2023Honolulu, HI
- XiongMHuZLuXLiYFuJHeJHooiBCan LLMs Express Their Uncertainty? An Empirical Evaluation of Confidence Elicitation in LLMsCan LLMs express their uncertainty? An empirical evaluation of confidence elicitation in LLMs2024The Twelfth International Conference on Learning RepresentationsMay 7, 2024Vienna, Austria
- GeifmanYEl-YanivRSelective classification for deep neural networks2017NIPS'17: Proceedings of the 31st International Conference on Neural Information Processing SystemsDecember 4-9, 2017Long Beach, CA48854894
- Authors
- Socrates, V., Wright, D. S., Huang, T., Fereydooni, S., Dien, C., Chi, L., Albano, J., Patterson, B., Sasidhar Kanaparthy, N., Wright, C. X., Loza, A., Chartash, D., Iscoe, M., & Taylor, R. A.
- Affiliations
- Yale University (US)
- Yale New Haven Hospital (US)
- University of Wisconsin–Madison (US)
- University College Dublin (IE)
- Corresponding author
- Taylor, Richard Andrew
- ORCID
- Venue
- JMIR Aging
- Publisher
- JMIR Publications Inc.
- Published
- 2025-04-11
- Volume
- 8
- Pages
- e69504
- ISSN
- 2561-7605 (ISSN-L)
- PMC
- PMC12032504
- OpenAlex
- W4407820636
- Semantic Scholar
- fb79926526ddd2208cb664a2844d7e81c859eb55
- Cited by
- 17
- References
- 38
- Citations by year
- 2026: 15 · 2025: 2
- FWCI
- 10.18
- Type
- journal article
- Language
- en
- Open access
- gold
- doi.org
- Licence
- cc-by
- Locations
- JMIR Aging (publishedVersion, cc-by)
- PubMed Central (publishedVersion, cc-by)
- PubMed (publishedVersion)
- DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
- Funders
- Yale University (UL1 TR001863)
- Georgia Clinical and Translational Science Alliance (UL1TR001863)
- National Institutes of Health (UL1TR001863)
- Center of Excellence in Regulatory Science and Innovation
- U.S. National Library of Medicine (T15LM007056)
- Topics
- Health Sciences › Medicine › Geriatrics and Gerontology › Pharmaceutical Practices and Patient Outcomes
- Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
- Physical Sciences › Computer Science › Artificial Intelligence › Machine Learning in Healthcare
- Keywords
- deprescribing, polypharmacy, older adults, emergency department, large language models, potentially inappropriate medications, Beers criteria, selective prediction, confidence calibration, clinical decision support, adverse drug reactions, falls, drug safety, LLM hallucination
- MeSH terms
- Potentially Inappropriate Medication List, Deprescriptions, Large Language Models, Aged, Aged, 80 and over, Emergency Service, Hospital, Female, Humans, Male, Retrospective Studies, Polypharmacy, Inappropriate Prescribing
About these checks
Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.
Claims read by deepseek-v4.1-flash under definition version 3 (extract-v3) on 2026-10-07.
Classified by jev-1.13-20260917 (gold-guideline-v0.2) on 2026-10-07.
Statistics and arithmetic are recomputed, not judged.
Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.
What counts as a claim ›How we measure confidence ›How we check this ›
Are you an author, or do you see a problem with this page? Request a review
