Recommending Clinical Trials for Online Patient Cases using Artificial Intelligence

Abstract

Clinical trials are crucial for assessing new treatments; however, recruitment challenges—such as limited awareness, complex eligibility criteria, and referral barriers—hinder their success. With the growth of online platforms, patients, caregivers, and family members increasingly post medical cases on social media and health communities, while physicians publish case reports accessible on platforms like PubMed—collectively expanding recruitment pools beyond traditional clinical trial pathways. Recognizing this potential, we utilized TrialGPT, a framework that leverages a large language model, to match 50 online patient cases (collected from case reports and social media) to clinical trials and evaluate performance against traditional keyword-based searches. Our results show that TrialGPT outperformed traditional methods by 46%, with patients eligible, on average, for 7 of the top 10 recommended trials. Additionally, outreach to case authors and trial organizers yielded positive feedback. These findings highlight TrialGPT’s potential to expand patient access to specialized care through non-traditional sources.

1. Introduction

Recruitment of patients to clinical trials remains a significant challenge that directly impacts trial success. Around 80% of clinical trials terminate due to the failure to meet enrollment targets within the expected timeframe, leading to wasted resources, delayed medical advancements, and limited treatment options for patients in need¹. The lack of participation from patients in clinical trials can be attributed to multiple factors, including limited awareness and understanding of clinical research, restrictive trial eligibility criteria, geographic and financial barriers, mistrust in medical institutions, concerns about potential risks and side effects, and low physician recommendations. However, the most significant barrier remains lack of awareness, as in 2020, only 9.4% of U.S. adults had full awareness of clinical trials, while 41% knew nothing about them^2,3.

Discovering suitable clinical trials is often complex, and patients rely on healthcare providers as their most trusted source of information. Getz (2017) found that 84% of patients would join a trial if referred by their physician^4,5. However, physicians face barriers such as limited time, insufficient trial information, and resource constraints, leading to missed enrollment opportunities. Consequently, many patients—especially those at facilities without active trials— must search for options independently. This can be overwhelming due to the complexity of medical information and eligibility criteria.

Increasingly, patients turn to online platforms like Reddit health communities and Facebook, where they connect with others facing similar conditions³⁰. Social media is particularly valuable for those with rare diseases, cancer, or chronic conditions, where standard treatments are often ineffective. Miller et al. found that many with rare diseases seek social media support due to limited treatment options¹⁸. Currently, 300 million people worldwide live with rare diseases, yet 95% lack treatments approved by the U.S. Food and Drug Administration²⁰. Similarly, 80% of cancer patients use social media for support and treatment exploration, with 49% citing diagnosis as their reason for joining, particularly in aggressive cases like glioblastoma^17,21. Those with chronic conditions also engage in peer discussions to navigate self-management and emerging treatments¹⁹. Beyond patient discussions, healthcare providers often share case reports on rare conditions and treatment outcomes^27,34. However, as patient cases are shared across multiple platforms in diverse formats, traditional clinical trial search methods struggle to keep pace, making it increasingly difficult to identify relevant trials for those outside standard treatment protocols.

We recently proposed TrialGPT⁸, which leverages a large language model (LLM) as its backbone to match patients with suitable clinical trials. TrialGPT, which has been evaluated with 183 synthetic patients across more than 75,000 trial eligibility annotations, demonstrated a reduced screening time by 42.6% in patient recruitment and high accuracy in patient-to-trial matching, achieving 87.3% accuracy in criterion-level eligibility predictions, closely aligning with expert performance.

To further validate its effectiveness with real-world data, we conducted an evaluation using 50 online patient cases sourced from published case reports in PubMed and three online health communities on Reddit. These cases were matched to clinical trials using TrialGPT and compared against results using the traditional keyword search method. In total, TrialGPT matched 500 patient-trial pairs, while the traditional keyword search method produced 475 pairs— yielding a combined total of 975 patient-trial pairs. The traditional keyword search method had 25 fewer matches because two online patient cases returned no clinical trials, and one case returned only 5. Then, two graduate-level researchers with biomedical knowledge manually evaluated the patient-trial pairs, and two internal clinicians selected the most promising clinical trials from these evaluations for outreach to the trial organizer and case authors— specifically for the PubMed case reports. This process details our study pipeline visualized in Figure 1. Our findings indicate that TrialGPT significantly outperformed traditional approaches, achieving an average eligibility rate of 75.7% compared to 29.7%. These results highlight TrialGPT’s potential to enhance clinical trial matching, particularly for patients with complex medical conditions.

Figure 1.

Study design. This figure illustrates the workflow for the curation and evaluation of clinical trial matches. a. Online patient cases published in 2024 on PubMed and posted on Reddit communities (AskDocs, Rare Diseases, and Cancer) were collected. b. The traditional keyword search method used the collected cases as input to prompt GPT-4o, which generated a complex Boolean query based on patient conditions. This query was then used in the traditional clinical trial search, resulting in a list of clinical trials for each patient. c. The collected cases were also used as input for TrialGPT, which consists of three components — retrieval, matching, and ranking — resulting in a ranked list clinical trials for each patient. d. Two graduate-level researchers with biomedical knowledge reviewed and validated the patient–trial pairs recommended by both TrialGPT and the traditional keyword search method. e. Both methods were compared using Precision, Mean Reciprocal Rank to assess eligibility, and Hit Rate to evaluate the identification of beneficial trials. f. Two internal clinicians, using their domain expertise, selected the most promising patient–trial pairs recommended by TrialGPT from the PubMed case reports for outreach. g. Case authors and trial organizers of the selected patient–trial pairs were contacted to assess match eligibility, trial benefit, and the clarity of TrialGPT explanations.

2. Methods

2.1. Datasets

Patient Summaries Data. To compose the 50 online patient cases, we collected patient summaries from various sources to ensure a diverse and representative dataset. We collected 25 recent patient summaries from case reports published on PubMed between January and August 2024, selecting cases involving individuals in the United States who exhibited ongoing symptoms following treatment. Additionally, we sourced 25 Reddit posts: 15 from the AskDocs community, 5 from a Cancer community, and 5 from a Rare Diseases community, all posted between May 2024 and August 2024. The dataset consists of retrospective, publicly available (published), de-identified clinical data and is therefore considered exempt from IRB review.

Table 1 provides an overview of the cases from different sources, highlighting key characteristics such as sex distribution, age distribution, and average case length. Notably, the patient summaries in our dataset are significantly longer than the synthetic summaries from TrialGPT’s original dataset—averaging approximately 469 words compared to 118 words, which were often brief. Moreover, while posts from Reddit communities tended to be verbose, the PubMed case reports provided more extensive and relevant medical information.

Table 1.

Statistics of online patient cases used in this study. Variables are reported as mean ± standard deviation where applicable. We report the number of cases (N), sex distribution, patient age, and case length (in words) across PubMed case reports and Reddit communities.

Data source	Case reports	Reddit – AskDocs	Reddit – Rare Diseases	Reddit – Cancer
N	25	15	5	5
Sex (male: female)	14: 11	4: 11	1: 4	2: 3
Age (year)	39 ± 23.6	33.4 ± 18.2	33.4 ± 19.0	34.4 ± 14.4
Case length (word)	602.0 ± 272.5	430.1 ± 142.5	319.8 ± 212.2	527 ± 208.3

Clinical Trials Data. We gathered data on approximately 23,000 actively recruiting clinical trials in the United States, including both interventional and observational studies, sourced from ClinicalTrials.gov (CT.gov) in July 2024. These trials were filtered solely by location, considering only those that are actively recruiting in the United States. Unlike many other countries that rely on multiple national or regional registries, the United States primarily uses CT.gov as its centralized database for registering and reporting clinical trials. This makes it the most comprehensive and consistent source for clinical trial data within a single country, ensuring data completeness and comparability. In contrast, trials conducted in other countries are often registered across multiple platforms, such as the WHO International Clinical Trials Registry Platform, EU Clinical Trials Register, or national registries, making it difficult to aggregate and analyze data systematically²⁹. For each trial, we collected its Brief Title, Phase, Drugs, Drugs List, Diseases, Diseases List, Enrollment, Inclusion Criteria, Exclusion Criteria, and Brief Summary.

2.2. TrialGPT

In this study, we leveraged GPT-4o (version 2024-02-01), accessed through Microsoft Azure’s OpenAI services, with the inference temperature fixed at 0 to ensure deterministic outputs. GPT-4o served as the backbone LLM for TrialGPT, which consists of three key components: TrialGPT-Retrieval, TrialGPT-Matching, and TrialGPT-Ranking. First, TrialGPT-Retrieval identifies relevant clinical trials using a combination of keyword generation and hybrid retrieval methods. The LLM extracts up to 32 relevant keywords from both the patient note and a summarized version of the note, which are then processed by a hybrid retrieval system that integrates both BM25³⁶ for lexical search and MedCPT⁹ for semantic search to retrieve the most relevant clinical trials. The top 1,000 trials from this process advance to the next stage. Next, TrialGPT-Matching performs a criterion-level analysis by comparing the patient summary to the inclusion and exclusion criteria of each retrieved trial. It identifies relevant sentences and provides explanations for each match. Finally, TrialGPT-Ranking orders the matched trials based on their scores. This ranking aggregates scores from TrialGPT-Matching’s criterion-by-criterion analysis, along with relevance and eligibility scores assigned by the LLM, ensuring that the most relevant trials are prioritized. In this study, the top 10 highest-ranked clinical trials are selected, resulting in a total of 500 patient-to-trial pairs. More details about TrialGPT are available in a prior publication⁸.

2.3. Evaluation

Evaluation Metrics. Our study evaluates patient-to-trial eligibility as the primary outcome and whether a trial is beneficial for patients as the secondary outcome, where a trial is evaluated for its benefit only if the patient is first deemed eligible for that trial. A patient is considered eligible if they meet all inclusion criteria and none of the exclusion criteria. A trial is deemed beneficial only if it provides a controlled treatment assessment and access to novel therapies that offer direct patient benefits. This includes the systematic evaluation of treatment efficacy, targeted therapies, or advanced medical strategies that are not yet available in standard clinical practice, thereby improving disease management and patient prognosis²⁹. We report eligibility using Precision for recommended trials at ranks 1, 3, 5, and 10, as well as the Mean Reciprocal Rank (MRR). Precision at rank k is defined as the proportion of eligible trials within the top k recommendations and can be computed as:

$$\text{P@k}\hspace{0.17em}=\hspace{0.17em}\frac{\text{Number}\hspace{0.17em}\text{of}\hspace{0.17em}\text{eligible}\hspace{0.17em}\text{trials}\hspace{0.17em}\text{in}\hspace{0.17em}\text{top}\hspace{0.17em}\text{k}}{k}$$ (1)

MRR measures how early the first eligible clinical trial appears in the ranked list and is defined as the average of the reciprocal ranks of the first relevant result for each patient case, which is computed as:

$$\text{MRR}\hspace{0.17em}=\hspace{0.17em}\frac{1}{N}{\displaystyle \sum _{i=1}^N\frac{1}{{\text{rank}}_i}}$$ (2)

where N is the total number of patient cases evaluated, and ranki is the position of the first eligible clinical trial in the ranked list for the i^th patient case.

We measured the Hit Rate at each trial rank, denoted by [t₁, t₂, …,t₁₀], by calculating, for each trial position, the proportion of patient cases for which a beneficial clinical trial was found at or before that specific trial rank. The Hit Rate at trial rank t is computed as:

$$\text{HitRate@t}\hspace{0.17em}=\hspace{0.17em}\frac{1}{N}{\displaystyle \sum _{i=1}^{N}1\left(\text{a}\hspace{0.17em}\text{beneficial}\hspace{0.17em}\text{trial}\hspace{0.17em}\text{found}\hspace{0.17em}\text{by}\hspace{0.17em}\text{rank}\hspace{0.17em}t\hspace{0.17em}\text{for}\hspace{0.17em}\text{patient}\hspace{0.17em}i\right)}$$ (3)

where N is the total number of patient cases evaluated, and 1(⋅) is an indicator function equal to 1 if a beneficial trial was found by trial rank t for patient i, and 0 otherwise.

These metrics comprehensively assess TrialGPT’s effectiveness in identifying relevant clinical trials, offering a quantitative measure of how efficiently and accurately the model can assist patients and clinicians in finding suitable trial opportunities.

Manual Evaluation. Two graduate-level researchers with biomedical knowledge evaluated the trials recommended by TrialGPT for the dataset of 50 online patient cases, resulting in 500 patient-trial matchings. One researcher assessed each patient summary against the eligibility criteria for each matched trial, determining whether the patient met all inclusion criteria and none of the exclusion criteria. If the trial was deemed eligible, the researcher then evaluated its potential benefit for the patient based on their medical condition. The second researcher validated these results to ensure agreement. The inter-rater agreement was 96.4%, and any disagreements were resolved through discussion.

Two internal clinicians were recruited to further evaluate and re-rank the eligible cases assessed by the two researchers on the PubMed case reports. They selected only the most promising and beneficial patient-trial pairs for outreach consideration. The clinicians were provided with patient summaries along with the clinical trials recommended by the researchers. They assessed whether the patient was eligible for the trial and whether the trial would provide meaningful medical benefits. Then, they determined whether outreach was necessary, which was defined as cases where the patient was both eligible and likely to benefit from enrollment, taking into account of additional considerations such as patient location and accessibility.

2.4. Outreach

For each patient-trial pair, both the case author — who is part of the healthcare team caring for the patient and presenting the case — and the trial organizers representing the clinical trials were contacted based on the clinicians’ outreach responses²³. Specifically, only the case authors of the PubMed patient summaries were contacted if clinicians deemed outreach necessary, while only the trial organizers of the corresponding trials were contacted when warranted. Both groups were engaged to gain a comprehensive perspective on TrialGPT’s performance. Contact information for case authors was obtained from the PubMed author information section, while trial organizers were identified through CT.gov. Their responses were recorded in a standardized form, with a follow-up sent one week after the initial contact. If no response was received within two weeks, the case was marked as unresponsive.

The outreach aimed to assess three key aspects: (1) patient-to-trial eligibility, (2) whether the clinical trial would helpful/benefit the patient (measured on a Likert scale from 1 to 5), and (3) the clarity of TrialGPT’s explanation for each criterion (also measured on a Likert scale from 1 to 5). The same set of questions was asked to both case authors and trial organizers to enable a direct comparison of their perspectives. Case authors were provided with the clinical trials that our internal clinicians had deemed eligible and beneficial, along with TrialGPT’s criterion-by-criterion analysis explaining the reasoning behind each match. Similarly, trial organizers received the online patient case along with TrialGPT’s reasoning, identical to what was provided to case authors but specific to their trial.

2.5. Comparing TrialGPT with Traditional Keyword Search Method

Traditional Keyword Search Method. We define traditional keyword search method as using CT.gov to identify relevant clinical trials. CT.gov is a database comprising over 400,000 registered studies¹⁴. Traditional keyword search consists of two approaches: basic and advanced. Since TrialGPT utilizes an LLM, we ensured a fair comparison by leveraging GPT-4o with zero-shot prompting to generate complex Boolean queries for the CT.gov conditions field, restricting results to actively recruiting trials in the United States as we had done for the clinical trials used in TrialGPT. This approach involved extracting relevant keywords from patient summaries, specifically focusing on relevant medical conditions.

Evaluation. Like our evaluation technique for TrialGPT, the same two researchers manually assessed the top 10 clinical trials retrieved through the traditional keyword search method for each online patient case, resulting in 475 patient-to-trial matches. The researchers applied the same evaluation criteria—eligibility and benefit—to the patient-trial pairs. The inter-rater agreement was 94.3%. To evaluate eligibility performance, we used Precision and Mean Reciprocal Rank to assess how accurately and effectively the system ranked eligible trials. For beneficial trials, we calculated the Hit Rate across the top 10 clinical trials recommended. This approach ensured consistency with the evaluation framework used for TrialGPT and provided a comprehensive assessment of clinical trial relevance.

3. Results

3.1. Overall Eligibility Performance

In this section, we compare the clinical trial recommendations by TrialGPT with those using the traditional keyword search method across the dataset, categorized by data source. A comparative analysis is presented in Table 2. We assessed patient eligibility for trials using Precision at ranks 1, 3, 5, and 10, where rank n refers to the top n recommended trials for a given patient. Additionally, we evaluated performance using the MRR. Upon examination, TrialGPT consistently outperformed the traditional keyword search method by a substantial margin across all metrics.

Table 2.

Eligibility Performance of TrialGPT and the Traditional Keyword Search Method. P@k refers to Precision at rank k, measuring the proportion of eligible trials within the top k recommendations. MRR indicates how highly the first eligible trial appears in the ranking, with higher values reflecting earlier correct matches. Values are reported per data source.

Method/Metric	Data source	P@1	P@3	P@5	P@10	MRR
Traditional Keyword Search	Case reports	0.440	0.266	0.232	0.208	0.520
	Reddit – AskDocs	0.333	0.312	0.320	0.280	0.457
	Reddit – Rare Diseases	0.400	0.467	0.360	0.500	0.600
	Reddit – Cancer	0.200	0.268	0.200	0.200	0.300
	Overall Average	0.343	0.328	0.278	0.297	0.469
TrialGPT	Case reports	0.880	0.850	0.840	0.820	0.930
	Reddit – AskDocs	0.725	0.746	0.736	0.748	0.800
	Reddit – Rare Diseases	0.800	0.934	0.920	0.760	0.900
	Reddit – Cancer	1.000	0.868	0.800	0.700	1.000
	Overall Average	0.851	0.849	0.824	0.757	0.975

More specifically, the average eligibility performance of TrialGPT surpasses that of the traditional keyword search method across all evaluated metrics. TrialGPT achieved an average improvement of 50.8% in Precision at rank 1 (P@1), 52.1% at rank 3 (P@3), 54.6% at rank 5 (P@5), and 46.0% at rank 10 (P@10). This means that patients are, on average, eligible for approximately 7 out of the top 10 recommended trials with TrialGPT, compared to only about 3 out of 10 with the traditional keyword search method. These results demonstrate a significant improvement over the traditional keyword search method, indicating that the process of identifying eligible clinical trials for patients is greatly enhanced with the use of TrialGPT. Additionally, TrialGPT showed a 50.1% improvement in MRR. While the traditional keyword search method typically presents the first eligible trial around the third result, TrialGPT, on average, ranks the first eligible trial at the very top of the list. This advantage is particularly crucial for patients with complex medical conditions, where timely access to appropriate clinical trials can significantly impact treatment decisions and outcomes. TrialGPT not only identifies more eligible trials but also ranks them higher, ensuring that patients and healthcare providers can quickly find the most relevant options without needing to sift through less pertinent results — a crucial advantage when every moment counts in guiding care and improving patient outcomes.

3.2. Evaluating Beneficial Outcomes in Top 10 Clinical Trials

In this section, we evaluate the performance of TrialGPT across the top 10 recommended clinical trials in comparison to the traditional keyword search method. Specifically, this analysis focuses on our secondary outcome: the proportion of beneficial trials identified among the top 10 trials for each patient. Performance was assessed using the Hit Rate, averaged across patients and stratified by data source. The results of this analysis are presented in Figure 2.

Figure 2.

Comparison of Hit Rate performance for identifying beneficial clinical trials using TrialGPT and the traditional keyword search method. a. Average beneficial Hit Rate across the top 10 recommended clinical trials for PubMed case reports. b. Average beneficial Hit Rate across the top 10 recommended clinical trials for Reddit communities (AskDocs, Rare Diseases, and Cancer).

Our findings indicate that TrialGPT is more effective than the traditional keyword search method in identifying beneficial clinical trials for patients across both data sources, as shown in Figure 2. Specifically, Figure 2a illustrates that TrialGPT outperformed the traditional keyword search method in identifying beneficial clinical trials from PubMed case reports, supporting our hypothesis that the presence of more precise medical terminology allows TrialGPT to perform better. In contrast, Figure 2b shows that the traditional keyword search method performed slightly better in identifying beneficial clinical trials from online patient cases from Reddit communities.

Overall, TrialGPT outperformed the traditional search method in identifying beneficial clinical trials across the entire dataset. When evaluating trial-by-trial performance for the PubMed case reports, we found that by the 10th recommended trial, TrialGPT achieved an average Hit Rate of 80%, compared to only 48% for the traditional keyword search method. In other words, patients using TrialGPT encountered a beneficial clinical trial within the first 10 recommendations 80% of the time, whereas those relying on the traditional keyword search method did so only 48% of the time. Similarly, in the Reddit posts, TrialGPT achieved a 68% Hit Rate, while the traditional search method reached only 52% by the 10th recommended trial. Although the difference in Reddit-based cases is smaller than in the PubMed case reports, this 16% improvement is substantial. It highlights that TrialGPT significantly increases the likelihood that patients — even those describing their conditions in informal, less structured language on platforms like Reddit — can quickly identify clinical trials that may have a direct impact on their care. This is especially important as an increasing number of patients turn to social media platforms and online health communities for medical guidance. Moreover, it is important to note that the recommended clinical trials from the traditional search method in this study relied on prompting an LLM to generate complex Boolean queries from patient summaries — a level of refinement that may not be practical or commonly used by patients or healthcare providers in real-world manual search scenarios. These findings demonstrate that TrialGPT is robust across both structured, provider-generated patient narratives and informal, patient-generated content, making it highly applicable to a wide range of real-world use cases.

3.3. Outreach Analysis

This section examines outreach responses from case authors and trial organizers contacted based on patient-trial pairs that clinicians identified as the most promising from TrialGPT’s clinical trial recommendations for the PubMed case reports. Outreach was conducted for 20 patient-trial pairs, with 20 trial organizers and 12 case authors contacted. More trial organizers were contacted because some patients matched multiple suitable clinical trials. Their perspectives on TrialGPT’s patient-to-trial matching recommendations offer insight into its role in clinical decision-making and trial selection interpretability. Although the response rate was low—8% for case authors and 15% for trial organizers—this can be attributed to several limitations. Nevertheless, these responses highlight the system’s practical impact from two key perspectives: healthcare providers and trial organizers.

Case Authors’ Perspective. Case authors who evaluated TrialGPT’s recommendations found their recommended trial or trials helpful and confirmed patient eligibility, noting that the explanations provided for each criterion were clear and informative. They also recognized TrialGPT’s ability to effectively explain its rationale, further reinforcing its usefulness in patient-to-trial matching.

Trial Organizers’ Perspective. Trial organizers also reviewed TrialGPT’s recommendations for different cases, providing insights into its role in refining patient enrollment decisions. While some found the recommended trials highly relevant and the eligibility explanations exceptionally clear, others, despite confirming patient eligibility, indicated that the trial’s overall helpfulness to the patient was low. In cases where a patient was not eligible for their trial, they still found the criterion explanations structured and informative.

Across both perspectives, TrialGPT received positive feedback for its ability to enhance interpretability and support decision-making in the trial selection process. Its detailed criterion-based explanations not only improved transparency but also helped both case authors and trial organizers better understand the rationale behind trial recommendations.

4. Related Works

4.1. Traditional Keyword Search for Clinical Trial Matching

The challenges of clinical trial matching have been widely studied in the past, with various approaches developed to improve the identification of eligible trials for patients. Traditional clinical trial search methods, including basic and advanced keyword search, often present challenges for both patients and healthcare providers. Basic keyword search returns numerous irrelevant results, requiring significant time to filter through, while advanced search relies on Boolean operators, demanding technical expertise. Even experienced users can spend over 100 hours refining search strategies^11,15. While these methods provide access to comprehensive trial databases, studies have shown that they are often ineffective at identifying suitable trials for patients without specialized knowledge.

4.2. Artificial Intelligence-Driven Clinical Trial Matching Systems

To improve efficiency and reduce manual review, studies have adopted Artificial Intelligence (AI) for clinical trial matching. Early rule-based systems used regex to filter ineligible patients but struggled with unstructured clinical notes due to variability and lack of standardization²⁴. To address this, ML and NLP techniques, including Named Entity Recognition (NER) and ontology-based normalization (SNOMED-CT, UMLS), were introduced. Hybrid rule-based and ML models improved accuracy but struggled with contextual understanding, negations, and temporal constraints²⁵. The introduction of LLMs has significantly improved semantic comprehension in trial matching³⁶. Key contributors include frameworks such as TrialGPT, PRISM, and Cognitive Technology for Cancer Trial Matching^26,27; systems like criteria2query³²; and models including DeepEnroll and COMPOSE^33,34. In this work, we demonstrate that TrialGPT outperforms the traditional keyword search method by achieving higher eligibility precision and identifying more beneficial trials, particularly for a dataset that is composed of real-world diverse online patient cases that haven’t ever been explored by clinical matching systems. Outreach results further highlight strong positive feedback from trial organizers and case authors, underscoring the promise of LLM-driven patient recruitment.

5. Discussion and Conclusions

In summary, this study successfully utilized TrialGPT to match online patient cases with relevant clinical trials. To the best of our knowledge, it is the first study to leverage such diverse online patient cases for trial matching, creating new opportunities for clinical trial recruitment that benefit both patients and healthcare providers. Our findings demonstrate that TrialGPT significantly outperforms the traditional keyword search method in identifying eligible trials, with improvements of 46%, and on average, patients are eligible for 7 of the top 10 recommended trials. This improvement is particularly evident for patients with rare diseases, cancer, and chronic conditions, where existing trial search mechanisms often fail to provide relevant matches.

Despite these promising results, several limitations must be acknowledged. First, the manual collection of online patient cases proved to be a time-intensive process, involving the identification, extraction, and review of patient narratives from public platforms. To enhance scalability, future work could incorporate LLMs to automate this step by using techniques such as entity recognition to identify and extract patient cases from large corpora of forum discussions. Additionally, case reports often exhibit publication bias, as authors are more likely to share rare conditions or positive treatment outcomes rather than persistent symptoms or treatment failures. This bias presents a challenge in capturing a comprehensive and representative view of real-world patient experiences.

Furthermore, although online patient cases often contained more detailed narratives compared to the synthetic patient cases used in TrialGPT evaluation, they frequently lacked critical medical details. This limitation made it difficult to fully assess whether a patient met all clinical trial eligibility criteria, particularly for trials with complex inclusion and exclusion requirements. Given the ethical considerations of this study, the inability to contact patients directly restricted opportunities to supplement missing information, further limiting the depth of clinical assessment. Future work could explore collaborations with healthcare institutions to facilitate follow-up with patients, allowing for a more comprehensive collection of clinical details while ensuring ethical compliance.

Another key limitation lies in the scope and scale of the online patient cases included in this study. Our evaluation was limited to 50 online patient cases, which restricts the generalizability of our findings. While these cases offer real-world diversity, larger datasets—particularly from standardized benchmarks or clinical settings—are needed to validate TrialGPT’s broader applicability. Additionally, we examined cases from only two platforms, PubMed and Reddit, and were unable to incorporate a wider range of patient communities such as PatientsLikeMe, Facebook, or X. Expanding to these platforms could provide a more comprehensive representation of patient narratives and shed light on how different online communities engage with clinical trial information. Because platforms attract distinct populations based on disease type, condition severity, and healthcare access, understanding these variations is essential for optimizing TrialGPT’s effectiveness across diverse groups. Future research should investigate engagement patterns across platforms and validate the system on larger, more representative datasets.

Additionally, our study focused on outreach to case authors and trial organizers to evaluate TrialGPT’s ability to determine patient eligibility and identify beneficial clinical trials. Consequently, only the 25 online patient cases from PubMed case reports were evaluated in detail by clinicians—who themselves often face heavy workloads. A low response rate from both groups made it challenging to comprehensively assess TrialGPT’s performance in a real-world clinical context. Case authors, often occupied with demanding clinical and research responsibilities, may not be the primary physician for the patient but rather a member of the care team. As a result, they might be less involved or may have already moved on, and the patient’s status might have changed. Additionally, it is uncommon for a corresponding author to receive such a request; they may have considered it spam or been less responsive, particularly since TrialGPT was applied to patient narratives written after its publication, making trial recommendations less of a priority. In contrast, trial organizers showed a slightly higher response rate, likely driven by their ongoing need for patient recruitment. These discrepancies highlight the importance of patients having timely access to clinical trial opportunities and the challenges they face when navigating the process independently. Many patients, especially those with rare diseases or complex conditions, struggle to identify suitable trials due to the complexity of eligibility criteria and limited guidance from healthcare providers who may not always be aware of available options or have the time to discuss these options with them.

In conclusion, this study highlights the feasibility and potential impact of using AI-driven models like TrialGPT to streamline clinical trial matching for a broader scope of patient cases—extending beyond traditional clinical notes or Electronic Health Records. Specifically, our focus on both clinician-derived patient narratives and patient-driven data demonstrates how these alternative sources of information can be leveraged to identify relevant trials. By automating patient-to-trial identification and providing transparent, criterion-level explanations for eligibility, TrialGPT serves as a valuable tool in bridging the gap between patients and clinical research. Moving forward, further validation and enhancements to TrialGPT will be crucial for optimizing its role in clinical trial recruitment, ultimately improving patient access to treatments and advancing medical research.