How suitable are clinical vignettes for the evaluation of symptom checker apps? A test theoretical perspective

Accuracy of online symptom assessment applications, large language models, and laypeople for self-triage decisions

Symptom-Assessment Application (SAAs, e.g., NHS 111 online) that assist laypeople in deciding if and where to seek care (self-triage) are gaining popularity and Large Language Models (LLMs) are increasingly used too. However, there is no evidence synthesis on the accuracy of LLMs, and no review has contextualized the accuracy of SAAs and LLMs. This systematic review evaluates the self-triage accuracy of both SAAs and LLMs and compares them to the accuracy of laypeople. A total of 1549 studies were screened and 19 included. The self-triage accuracy of SAAs was moderate but highly variable (11.5-90.0%), while the accuracy of LLMs (57.8-76.0%) and laypeople (47.3-62.4%) was moderate with low variability. Based on the available evidence, the use of SAAs or LLMs should neither be universally recommended nor discouraged; rather, we suggest that their utility should be assessed based on the specific use case and user group under consideration.

The RepVig framework for designing use-case specific representative vignettes and evaluating triage accuracy of laypeople and symptom assessment applications

Most studies evaluating symptom-assessment applications (SAAs) rely on a common set of case vignettes that are authored by clinicians and devoid of context, which may be representative of clinical settings but not of situations where patients use SAAs. Assuming the use case of self-triage, we used representative design principles to sample case vignettes from online platforms where patients describe their symptoms to obtain professional advice and compared triage performance of laypeople, SAAs (e.g., WebMD or NHS 111), and Large Language Models (LLMs, e.g., GPT-4 or Claude) on representative versus standard vignettes. We found performance differences in all three groups depending on vignette type: When using representative vignettes, accuracy was higher (OR = 1.52 to 2.00, p < .001 to .03 in binary decisions, i.e., correct or incorrect), safety was higher (OR = 1.81 to 3.41, p < .001 to .002 in binary decisions, i.e., safe or unsafe), and the inclination to overtriage was also higher (OR = 1.80 to 2.66, p < .001 to p = .035 in binary decisions, overtriage or undertriage error). Additionally, we found changed rankings of best-performing SAAs and LLMs. Based on these results, we argue that our representative vignette sampling approach (that we call the RepVig Framework) should replace the practice of using a fixed vignette set as standard for SAA evaluation studies.

Statistical refinement of patient-centered case vignettes for digital health research

Digital health research often relies on case vignettes (descriptions of fictitious or real patients) to navigate ethical and practical challenges. Despite their utility, the quality and lack of standardization of these vignettes has often been criticized, especially in studies on symptom-assessment applications (SAAs) and self-triage decision-making. To address this, our paper introduces a method to refine an existing set of vignettes, drawing on principles from classical test theory. First, we removed any vignette with an item difficulty of zero and an item-total correlation below zero. Second, we stratified the remaining vignettes to reflect the natural base rates of symptoms that SAAs are typically approached with, selecting those vignettes with the highest item-total correlation in each quota. Although this two-step procedure reduced the size of the original vignette set by 40%, comparing self-triage performance on the reduced and the original vignette sets, we found a strong correlation (r = 0.747 to r = 0.997, p < .001). This indicates that using our refinement method helps identifying vignettes with high predictive power of an agent's self-triage performance while simultaneously increasing cost-efficiency of vignette-based evaluation studies. This might ultimately lead to higher research quality and more reliable results.

Triage Accuracy and the Safety of User-Initiated Symptom Assessment With an Electronic Symptom Checker in a Real-Life Setting: Instrument Validation Study

BACKGROUND

Previous studies have evaluated the accuracy of the diagnostics of electronic symptom checkers (ESCs) and triage using clinical case vignettes. National Omaolo digital services (Omaolo) in Finland consist of an ESC for various symptoms. Omaolo is a medical device with a Conformité Européenne marking (risk class: IIa), based on Duodecim Clinical Decision Support, EBMEDS.

OBJECTIVE

This study investigates how well triage performed by the ESC nurse triage within the chief symptom list available in Omaolo (anal region symptoms, cough, diarrhea, discharge from the eye or watery or reddish eye, headache, heartburn, knee symptom or injury, lower back pain or injury, oral health, painful or blocked ear, respiratory tract infection, sexually transmitted disease, shoulder pain or stiffness or injury, sore throat or throat symptom, and urinary tract infection). In addition, the accuracy, specificity, sensitivity, and safety of the Omaolo ESC were assessed.

METHODS

This is a clinical validation study in a real-life setting performed at multiple primary health care (PHC) centers across Finland. The included units were of the walk-in model of primary care, where no previous phone call or contact was required. Upon arriving at the PHC center, users (patients) answered the ESC questions and received a triage recommendation; a nurse then assessed their triage. Findings on 877 patients were analyzed by matching the ESC recommendations with triage by the triage nurse.

RESULTS

Safe assessments by the ESC accounted for 97.6% (856/877; 95% CI 95.6%-98.0%) of all assessments made. The mean of the exact match for all symptom assessments was 53.7% (471/877; 95% CI 49.2%-55.9%). The mean value of the exact match or overly conservative but suitable for all (ESC's assessment was 1 triage level higher than the nurse's triage) symptom assessments was 66.6% (584/877; 95% CI 63.4%-69.7%). When the nurse concluded that urgent treatment was needed, the ESC's exactly matched accuracy was 70.9% (244/344; 95% CI 65.8%-75.7%). Sensitivity for the Omaolo ESC was 62.6% and specificity of 69.2%. A total of 21 critical assessments were identified for further analysis: there was no indication of compromised patient safety.

CONCLUSIONS

The primary objectives of this study were to evaluate the safety and to explore the accuracy, specificity, and sensitivity of the Omaolo ESC. The results indicate that the ESC is safe in a real-life setting when appraised with assessments conducted by triage nurses. Furthermore, the Omaolo ESC exhibits the potential to guide patients to appropriate triage destinations effectively, helping them to receive timely and suitable care.

INTERNATIONAL REGISTERED REPORT IDENTIFIER (IRRID)

RR2-10.2196/41423.

Controlling Inputter Variability in Vignette Studies Assessing Web-Based Symptom Checkers: Evaluation of Current Practice and Recommendations for Isolated Accuracy Metrics

BACKGROUND

The rapid growth of web-based symptom checkers (SCs) is not matched by advances in quality assurance. Currently, there are no widely accepted criteria assessing SCs' performance. Vignette studies are widely used to evaluate SCs, measuring the accuracy of outcome. Accuracy behaves as a composite metric as it is affected by a number of individual SC- and tester-dependent factors. In contrast to clinical studies, vignette studies have a small number of testers. Hence, measuring accuracy alone in vignette studies may not provide a reliable assessment of performance due to tester variability.

OBJECTIVE

This study aims to investigate the impact of tester variability on the accuracy of outcome of SCs, using clinical vignettes. It further aims to investigate the feasibility of measuring isolated aspects of performance.

METHODS

Healthily's SC was assessed using 114 vignettes by 3 groups of 3 testers who processed vignettes with different instructions: free interpretation of vignettes (free testers), specified chief complaints (partially free testers), and specified chief complaints with strict instruction for answering additional symptoms (restricted testers). κ statistics were calculated to assess agreement of top outcome condition and recommended triage. Crude and adjusted accuracy was measured against a gold standard. Adjusted accuracy was calculated using only results of consultations identical to the vignette, following a review and selection process. A feasibility study for assessing symptom comprehension of SCs was performed using different variations of 51 chief complaints across 3 SCs.

RESULTS

Intertester agreement of most likely condition and triage was, respectively, 0.49 and 0.51 for the free tester group, 0.66 and 0.66 for the partially free group, and 0.72 and 0.71 for the restricted group. For the restricted group, accuracy ranged from 43.9% to 57% for individual testers, averaging 50.6% (SD 5.35%). Adjusted accuracy was 56.1%. Assessing symptom comprehension was feasible for all 3 SCs. Comprehension scores ranged from 52.9% and 68%.

CONCLUSIONS

We demonstrated that by improving standardization of the vignette testing process, there is a significant improvement in the agreement of outcome between testers. However, significant variability remained due to uncontrollable tester-dependent factors, reflected by varying outcome accuracy. Tester-dependent factors, combined with a small number of testers, limit the reliability and generalizability of outcome accuracy when used as a composite measure in vignette studies. Measuring and reporting different aspects of SC performance in isolation provides a more reliable assessment of SC performance. We developed an adjusted accuracy measure using a review and selection process to assess data algorithm quality. In addition, we demonstrated that symptom comprehension with different input methods can be feasibly compared. Future studies reporting accuracy need to apply vignette testing standardization and isolated metrics.

[Short paths to diagnosis with artificial intelligence: systematic literature review on diagnostic decision support systems]

BACKGROUND

Rare diseases are often recognized late. Their diagnosis is particularly challenging due to the diversity, complexity and heterogeneity of clinical symptoms. Computer-aided diagnostic aids, often referred to as diagnostic decision support systems (DDSS), are promising tools for shortening the time to diagnosis. Despite initial positive evaluations, DDSS are not yet widely used, partly due to a lack of integration with existing clinical or practice information systems.

OBJECTIVE

This article provides an insight into currently existing diagnostic support systems that function without access to electronic patient records and only require information that is easily obtainable.

MATERIALS AND METHODS

A systematic literature search identified eight articles on DDSS that can assist in the diagnosis of rare diseases with no need for access to electronic patient records or other information systems in practices and hospitals. The main advantages and disadvantages of the identified rare disease diagnostic support systems were extracted and summarized.

RESULTS

Symptom checkers and DDSS based on portrait photos and pain drawings already exist. The degree of maturity of these applications varies.

CONCLUSION

DDSS currently still face a number of challenges, such as concerns about data protection and accuracy, and acceptance and awareness continue to be rather low. On the other hand, there is great potential for faster diagnosis, especially for rare diseases, which are easily overlooked due to their large number and the low awareness of them. The use of DDSS should therefore be carefully considered by doctors on a case-by-case basis.