Current safeguards, risk mitigation, and transparency measures of large language models against the generation of health disinformation: repeated cross sectional analysis.

OBJECTIVES: To evaluate the effectiveness of safeguards to prevent large language models (LLMs) from being misused to generate health disinformation, and to evaluate the transparency of artificial intelligence (AI) developers regarding their risk mitigation processes against observed vulnerabilities. DESIGN: Repeated cross sectional analysis. SETTING: Publicly accessible LLMs. METHODS: In a repeated cross sectional analysis, four LLMs (via chatbots/assistant interfaces) were evaluated: OpenAI's GPT-4 (via ChatGPT and Microsoft's Copilot), Google's PaLM 2 and newly released Gemini Pro (via Bard), Anthropic's Claude 2 (via Poe), and Meta's Llama 2 (via HuggingChat). In September 2023, these LLMs were prompted to generate health disinformation on two topics: sunscreen as a cause of skin cancer and the alkaline diet as a cancer cure. Jailbreaking techniques (ie, attempts to bypass safeguards) were evaluated if required. For LLMs with observed safeguarding vulnerabilities, the processes for reporting outputs of concern were audited. 12 weeks after initial investigations, the disinformation generation capabilities of the LLMs were re-evaluated to assess any subsequent improvements in safeguards. MAIN OUTCOME MEASURES: The main outcome measures were whether safeguards prevented the generation of health disinformation, and the transparency of risk mitigation processes against health disinformation. RESULTS: Claude 2 (via Poe) declined 130 prompts submitted across the two study timepoints requesting the generation of content claiming that sunscreen causes skin cancer or that the alkaline diet is a cure for cancer, even with jailbreaking attempts. GPT-4 (via Copilot) initially refused to generate health disinformation, even with jailbreaking attempts-although this was not the case at 12 weeks. In contrast, GPT-4 (via ChatGPT), PaLM 2/Gemini Pro (via Bard), and Llama 2 (via HuggingChat) consistently generated health disinformation blogs. In September 2023 evaluations, these LLMs facilitated the generation of 113 unique cancer disinformation blogs, totalling more than 40 000 words, without requiring jailbreaking attempts. The refusal rate across the evaluation timepoints for these LLMs was only 5% (7 of 150), and as prompted the LLM generated blogs incorporated attention grabbing titles, authentic looking (fake or fictional) references, fabricated testimonials from patients and clinicians, and they targeted diverse demographic groups. Although each LLM evaluated had mechanisms to report observed outputs of concern, the developers did not respond when observations of vulnerabilities were reported. CONCLUSIONS: This study found that although effective safeguards are feasible to prevent LLMs from being misused to generate health disinformation, they were inconsistently implemented. Furthermore, effective processes for reporting safeguard problems were lacking. Enhanced regulation, transparency, and routine auditing are required to help prevent LLMs from contributing to the mass generation of health disinformation.

Accessibility of clinical study reports supporting medicine approvals: a cross-sectional evaluation.

OBJECTIVES: Clinical study reports (CSRs) are highly detailed documents that play a pivotal role in medicine approval processes. Though not historically publicly available, in recent years, major entities including the European Medicines Agency (EMA), Health Canada, and the US Food and Drug Administration (FDA) have highlighted the importance of CSR accessibility. The primary objective herein was to determine the proportion of CSRs that support medicine approvals available for public download as well as the proportion eligible for independent researcher request via the study sponsor. STUDY DESIGN AND SETTING: This cross-sectional study examined the accessibility of CSRs from industry-sponsored clinical trials whose results were reported in the FDA-authorized drug labels of the top 30 highest-revenue medicines of 2021. We determined (1) whether the CSRs were available for download from a public repository, and (2) whether the CSRs were eligible for request by independent researchers based on trial sponsors' data sharing policies. RESULTS: There were 316 industry-sponsored clinical trials with results presented in the FDA-authorized drug labels of the 30 sampled medicines. Of these trials, CSRs were available for public download from 70 (22%), with 37 available at EMA and 40 at Health Canada repositories. While pharmaceutical company platforms offered no direct downloads of CSRs, sponsors confirmed that CSRs from 183 (58%) of the 316 clinical trials were eligible for independent researcher request via the submission of a research proposal. Overall, 218 (69%) of the sampled clinical trials had CSRs available for public download and/or were eligible for request from the trial sponsor. CONCLUSION: CSRs were available from 69% of the clinical trials supporting regulatory approval of the 30 medicines sampled. However, only 22% of the CSRs were directly downloadable from regulatory agencies, the remaining required a formal application process to request access to the CSR from the study sponsor.

Health Disinformation Use Case Highlighting the Urgent Need for Artificial Intelligence Vigilance: Weapons of Mass Disinformation.

Importance: Although artificial intelligence (AI) offers many promises across modern medicine, it may carry a significant risk for the mass generation of targeted health disinformation. This poses an urgent threat toward public health initiatives and calls for rapid attention by health care professionals, AI developers, and regulators to ensure public safety. Observations: As an example, using a single publicly available large-language model, within 65 minutes, 102 distinct blog articles were generated that contained more than 17 000 words of disinformation related to vaccines and vaping. Each post was coercive and targeted at diverse societal groups, including young adults, young parents, older persons, pregnant people, and those with chronic health conditions. The blogs included fake patient and clinician testimonials and obeyed prompting for the inclusion of scientific-looking referencing. Additional generative AI tools created an accompanying 20 realistic images in less than 2 minutes. This process was undertaken by health care professionals and researchers with no specialized knowledge in bypassing AI guardrails, relying solely on publicly available information. Conclusions and Relevance: These observations demonstrate that when the guardrails of AI tools are insufficient, the ability to rapidly generate diverse and large amounts of convincing disinformation is profound. Beyond providing 2 example scenarios, these findings demonstrate an urgent need for robust AI vigilance. The AI tools are rapidly progressing; alongside these advancements, emergent risks are becoming increasingly apparent. Key pillars of pharmacovigilance-including transparency, surveillance, and regulation-may serve as valuable examples for managing these risks and safeguarding public health.

A 10-year update to the principles for clinical trial data sharing by pharmaceutical companies: perspectives based on a decade of literature and policies.

Data sharing is essential for promoting scientific discoveries and informed decision-making in clinical practice. In 2013, PhRMA/EFPIA recognised the importance of data sharing and supported initiatives to enhance clinical trial data transparency and promote scientific advancements. However, despite these commitments, recent investigations indicate significant scope for improvements in data sharing by the pharmaceutical industry. Drawing on a decade of literature and policy developments, this article presents perspectives from a multidisciplinary team of researchers, clinicians, and consumers. The focus is on policy and process updates to the PhRMA/EFPIA 2013 data sharing commitments, aiming to enhance the sharing and accessibility of participant-level data, clinical study reports, protocols, statistical analysis plans, lay summaries, and result publications from pharmaceutical industry-sponsored trials. The proposed updates provide clear recommendations regarding which data should be shared, when it should be shared, and under what conditions. The suggested improvements aim to develop a data sharing ecosystem that supports science and patient-centred care. Good data sharing principles require resources, time, and commitment. Notwithstanding these challenges, enhancing data sharing is necessary for efficient resource utilization, increased scientific collaboration, and better decision-making for patients and healthcare professionals.

Heterogeneity and Utility of Pharmaceutical Company Sharing of Individual-Participant Data Packages.

Importance: The pharmaceutical industry has made substantial investments in developing processes for sharing individual-participant data (IPD) from clinical trials. However, the utility and completeness of shared IPD and supporting documents must be evaluated to ensure the potential for scientific advancements from the data sharing ecosystem can be realized. Objective: To assess the utility and completeness of IPD and supporting documents provided from industry-sponsored clinical trials. Design, Setting, and Participants: From February 9, 2022, to February 9, 2023, 91 of 203 clinical trials supporting US Food and Drug Administration registrations of anticancer medicines for the treatment of solid tumors from the past decade were confirmed as eligible for IPD request. This quality improvement study performed a retrospective audit of the utility and completeness of the IPD and supporting documents provided from the 91 clinical trials for a planned meta-analysis. Exposures: Request for IPD from 91 clinical oncology trials indicated as eligible for the request. Main Outcomes and Measures: The utility and completeness of the IPD and supporting documents provided. Results: The IPD packages were obtained from 70 of 91 requested clinical trials (77%). The median time to data provision was 123 (range, 117-352) days. Redactions were observed in 18 of the acquired IPD packages (26%) for outcome data, 11 (16%) for assessment variables, and 19 (27%) for adjustment data. Additionally, 20 IPD packages (29%) lacked a clinical study report, 4 (6%) had incomplete or missing data dictionaries, and 20 (29%) were missing anonymization or redaction description files. Access to IPD from 21 eligible trials (23%) was not granted. Conclusions and Relevance: In this quality improvement study, there was substantial variability within the provided IPD packages regarding the completeness of key data variables and supporting documents. To improve the data sharing ecosystem, key areas for enhancement include (1) ensuring that clinical trials are eligible for IPD sharing, (2) making eligible IPD transparently accessible, and (3) ensuring that IPD packages meet a standard of utility and completeness.

Audit of Data Sharing by Pharmaceutical Companies for Anticancer Medicines Approved by the US Food and Drug Administration.

Importance: Emerging policies drafted by the pharmaceutical industry indicate that they will transparently share clinical trial data. These data offer an unparalleled opportunity to advance evidence-based medicine and support decision-making. Objective: To evaluate the eligibility of independent, qualified researchers to access individual participant data (IPD) from oncology trials that supported US Food and Drug Administration (FDA) approval of new anticancer medicines within the past 10 years. Design, Setting, and Participants: In this quality improvement study, a cross-sectional analysis was performed of pivotal clinical trials whose results supported FDA-approved anticancer medicines between January 1, 2011, and June 30, 2021. These trials' results were identified from product labels. Exposures: Eligibility for IPD sharing was confirmed by identification of a public listing of the trial as eligible for sharing or by receipt of a positive response from the sponsor to a standardized inquiry. Main Outcomes and Measures: The main outcome was frequency of IPD sharing eligibility. Reasons for data sharing ineligibility were requested and collated, and company-, drug-, and trial-level subgroups were evaluated and presented using χ2 tests and forest plots. Results: During the 10-year period examined, 115 anticancer medicines were approved by the FDA on the basis of evidence from 304 pharmaceutical industry-sponsored trials. Of these trials, 136 (45%) were eligible for IPD sharing and 168 (55%) were not. Data sharing rates differed substantially among industry sponsors, with the most common reason for not sharing trial IPD being that the collection of long-term follow-up data was still ongoing (89 of 168 trials [53%]). Of the top 10 anticancer medicines by global sales, nivolumab, pembrolizumab, and pomalidomide had the lowest eligibility rates for data sharing (<10% of trials). Conclusions and Relevance: There has been a substantial increase in IPD sharing for industry-sponsored oncology trials over the past 5 years. However, this quality improvement study found that more than 50% of queried trials for FDA-approved anticancer medicines were ineligible for IPD sharing. Data accessibility would be substantially improved if, at the time of FDA registration of a medicine, all data that support the registration were made available.

Prediction of severe neutropenia and diarrhoea in breast cancer patients treated with abemaciclib.

INTRODUCTION: Neutropenia and diarrhoea are common and potentially serious adverse events associated with abemaciclib in advanced breast cancer (ABC), and the risk factors have been minimally explored. The study aimed to develop clinical prediction tools that allow personalized predictions of neutropenia and diarrhoea following abemaciclib initiation. MATERIALS AND METHODS: Data was pooled from MONARCH 1, 2 and 3 trials investigating abemaciclib. Cox proportional hazard analysis was used to assess the association between pre-treatment clinicopathological data and grade ≥3 diarrhoea and neutropenia occurring within the first 365 days of abemaciclib use. RESULTS: Older age was associated with increased risk of grade ≥3 diarrhoea [HR [95%CI] for age > 70: 1.72 [1.14-2.58]; P = 0.009]. A clinical prediction tool for abemaciclib induced grade ≥3 neutropenia was optimally defined by race, ECOGPS and white blood cell count. Large discrimination between subgroups was observed; the highest risk subgroup had a 64% probability of grade ≥3 neutropenia within the first 365 days of abemaciclib (150 mg twice daily) + fulvestrant/NSAI, compared to 5% for the lowest risk subgroup. CONCLUSION: The study identified advanced age as significantly associated with an increased risk of abemaciclib induced grade ≥ 3 diarrhoea. A clinical prediction tool, defined by race, ECOGPS and pre-treatment white blood cell count, was able to discriminate subgroups with significantly different risks of grade ≥3 neutropenia following abemaciclib initiation. The tool may enable improved interpretation of personalized risks and the risk-benefit ratio of abemaciclib.

The obesity paradox in early and advanced HER2 positive breast cancer: pooled analysis of clinical trial data.

While many studies have evaluated the relationship between BMI and breast cancer outcomes, it is unclear whether this relationship is consistent between early breast cancer (BC) and advanced BC. The study included 5099 patients with HER2 positive early BC (EBC) and 3496 with HER2 positive advanced BC (ABC). In the EBC cohort, higher BMI was associated with worse overall survival (OS) (HR [95% CI]: overweight = 1.30 [1.13-1.51]; obese = 1.37 [1.14-1.64], P = < 0.001), and worse disease-free survival (overweight = 1.10 [0.98-1.24]; obese = 1.20 [1.04-1.39], P = 0.061). In contrast, for the ABC cohort, higher BMI was significantly associated with improved OS (overweight = 0.85 [0.76-0.96]; obese = 0.82 [0.72-0.95], P = 0.014), and progression-free survival (overweight = 0.91 [0.83-1.01]; obese = 0.87 [0.77-0.98], P = 0.034). In this large high-quality dataset, higher BMI was independently associated with worse survival in EBC, paradoxically in ABC higher BMI was independently associated with improved survival.

The Influence of Pre-Existing Beta-Blockers Use on Survival Outcomes in HER2 Positive Advanced Breast Cancer: Pooled Analysis of Clinical Trial Data.

Introduction: Beta-blockers (BB) are commonly used to manage cardiovascular disease and may have benefits in controlling complications of anti-HER2 therapies. This study aimed to evaluate the association of pre-existing BB use with survival outcomes in patients initiating anti-HER2 therapy for advanced breast cancer (ABC). Materials and Methods: Data from clinical trials EMILIA, TH3RESA, MARIANNE, and CLEOPATRA was pooled. Cox proportional analysis was used to assess the association between pre-existing BB use with survival outcomes in patients initiating anti-HER2 therapies. Results: Of the 2,777 patients with HER2 positive ABC, 266 were using a BB at the time of anti-HER2 therapy initiation. BB use was associated with worse overall survival (OS) (adjusted HR = 1.27, 95% CI: 1.04-1.55). Sensitivity analysis in patients with pre-existing cardiovascular disease (CVD) also indicated that BB use was associated with worse OS (1.29, 1.02-1.63). Conclusion: In large high-quality data, BB use at the time of anti-HER2 therapy initiation for ABC was independently associated with worse OS, regardless of CVD status. The finding is contrary to pre-study hypotheses and findings in other BC subtypes. Future research should aim to gain a deeper understanding of the effects of BBs on specific BC subtypes, cancer types, and cancer treatments.

Early Adverse Events predict Survival Outcomes in HER2-positive Advanced Breast Cancer Patients treated with Lapatinib plus Capecitabine.

Background: This study aimed to investigate the impact of early adverse events (AE) following the initiation of lapatinib plus capecitabine on the progression-free survival (PFS) and overall survival (OS) outcomes of human epidermal growth factor receptor 2 (HER2) positive advanced breast cancer (ABC) patients. Methods: A secondary analysis of participants treated with lapatinib plus capecitabine, or ado-trastuzumab emtansine in the clinical trial EMILIA was conducted. Cox proportional hazard analysis was used to assess the impact of AE occurring within the first 42 days of lapatinib plus capecitabine therapy on the PFS and OS outcomes of ABC patients. Results: The study included 488 HER2-positive (ABC) patients initiated on lapatinib plus capecitabine. Rash (Hazard Ratio (HR) [95% Confidence Interval (CI)]: Grade 1 = 0.67 [0.46-0.98], Grade 2+ = 0.71 [0.42-1.19]; p = 0.046) and hand-foot syndrome (HR [95% CI]: Grade 1 = 0.58 [0.43-0.80], Grade 2+ = 0.61 [0.43-0.86]; p = <0.001) occurring within the first 42 days of lapatinib plus capecitabine therapy were significantly associated with improved OS. Conversely, nausea and vomiting occurring within the first 42 days of lapatinib plus capecitabine therapy was significantly associated with worsened OS (HR [95% CI]: Grade 1 = 1.08 [0.82-1.42], Grade 2+ = 1.52 [1.13-2.03]; p = 0.027). Conclusions: Rash and hand-foot syndrome occurring early after the initiation of on lapatinib plus capecitabine were significantly associated with improved OS, while early nausea and vomiting was associated with worse OS. In HER2-positive ABC patients initiating lapatinib plus capecitabine, consideration should be given to more closely monitoring patients at risk of nausea and vomiting, while rash and hand foot syndrome are AE associated with improved survival.

A literature review of treatment-specific clinical prediction models in patients with breast cancer.

Despite advances in the breast cancer treatment, significant variability in patient outcomes remain. This results in significant stress to patients and clinicians. Treatment-specific clinical prediction models allow patients to be matched against historical outcomes of patients with similar characteristics; thereby reducing uncertainty by providing personalised estimates of benefits, harms, and prognosis. To achieve this objective, models need to be clinical-grade with evidence of accuracy, reproducibility, generalizability, and be user-friendly. A structured search was undertaken to identify treatment-specific clinical prediction models for therapeutic or adverse outcomes in breast cancer using clinicopathological data. Significant gaps in the presence of validated models for available treatments was identified, along with gaps in prediction of therapeutic and adverse outcomes. Most models did not have user-friendly tools available. With the aim being to facilitate the selection of the best medicine for a specific patient and shared-decision making, future research will need to address these gaps.

Predicting Thrombocytopenia in Patients With Breast Cancer Treated With Ado-trastuzumab Emtansine.

INTRODUCTION: Thrombocytopenia is a common and potentially serious adverse event of ado-trastuzumab emtansine (T-DM1) use in patients with advanced breast cancer. However, the risk factors have been minimally explored. Our aim was to develop a clinical prediction model from the clinicopathologic data that would allow for quantification of the personalized risks of thrombocytopenia from T-DM1 usage. MATERIALS AND METHODS: Data from 3 clinical trials, EMILIA (a study of trastuzumab emtansine versus capecitabine + lapatinib in participants with HER2 [human epidermal growth factor receptor 2]-positive locally advanced or metastatic breast cancer), TH3RESA (a study of trastuzumab emtansine in comparison with treatment of physician's choice in participants with HER2-positive breast cancer who have received at least two prior regimens of HER2-directed therapy), and MARIANNE [a study of trastuzumab emtansine (T-DM1) plus pertuzumab/pertuzumab placebo versus trastuzumab (Herceptin) plus a taxane in participants with metastatic breast cancer], were pooled. Cox proportional hazard analysis was used to assess the association between the pretreatment clinicopathologic data and grade ≥ 3 thrombocytopenia occurring within the first 365 days of T-DM1 use. A multivariable clinical prediction model was developed using a backward elimination process. RESULTS: Of the 1620 participants, 141 (9%) had experienced grade ≥ 3 thrombocytopenia. On univariable analysis, the body mass index, race, presence of brain metastasis, platelet count, white blood cell count, and concomitant corticosteroid use were significantly associated with the occurrence of grade ≥ 3 thrombocytopenia (P < .05). The multivariable prediction model was optimally defined by race (Asian vs. non-Asian) and platelet count (100-220 vs. 220-300 vs. >300 × 109/L). A large discrimination between the prognostic subgroups was observed. The highest risk subgroup (Asian and platelet count of 100-220 cells ×109/L) had a 40% probability of grade ≥ 3 thrombocytopenia within the first 365 days of T-DM1 therapy compared with 2% for the lowest risk subgroup (non-Asian and platelet count > 300 × 109/L). CONCLUSION: A clinical prediction model, defined by race and pretreatment platelet count, was able to discriminate subgroups with a significantly different risk of grade ≥ 3 thrombocytopenia after T-DM1 initiation. The model allows for improved interpretation of the personalized risks and risk/benefit ratio of T-DM1.