Large language models encode clinical knowledge.

Large language models (LLMs) have demonstrated impressive capabilities, but the bar for clinical applications is high. Attempts to assess the clinical knowledge of models typically rely on automated evaluations based on limited benchmarks. Here, to address these limitations, we present MultiMedQA, a benchmark combining six existing medical question answering datasets spanning professional medicine, research and consumer queries and a new dataset of medical questions searched online, HealthSearchQA. We propose a human evaluation framework for model answers along multiple axes including factuality, comprehension, reasoning, possible harm and bias. In addition, we evaluate Pathways Language Model1 (PaLM, a 540-billion parameter LLM) and its instruction-tuned variant, Flan-PaLM2 on MultiMedQA. Using a combination of prompting strategies, Flan-PaLM achieves state-of-the-art accuracy on every MultiMedQA multiple-choice dataset (MedQA3, MedMCQA4, PubMedQA5 and Measuring Massive Multitask Language Understanding (MMLU) clinical topics6), including 67.6% accuracy on MedQA (US Medical Licensing Exam-style questions), surpassing the prior state of the art by more than 17%. However, human evaluation reveals key gaps. To resolve this, we introduce instruction prompt tuning, a parameter-efficient approach for aligning LLMs to new domains using a few exemplars. The resulting model, Med-PaLM, performs encouragingly, but remains inferior to clinicians. We show that comprehension, knowledge recall and reasoning improve with model scale and instruction prompt tuning, suggesting the potential utility of LLMs in medicine. Our human evaluations reveal limitations of today's models, reinforcing the importance of both evaluation frameworks and method development in creating safe, helpful LLMs for clinical applications.

Non-viral precision T cell receptor replacement for personalized cell therapy.

T cell receptors (TCRs) enable T cells to specifically recognize mutations in cancer cells1-3. Here we developed a clinical-grade approach based on CRISPR-Cas9 non-viral precision genome-editing to simultaneously knockout the two endogenous TCR genes TRAC (which encodes TCRα) and TRBC (which encodes TCRß). We also inserted into the TRAC locus two chains of a neoantigen-specific TCR (neoTCR) isolated from circulating T cells of patients. The neoTCRs were isolated using a personalized library of soluble predicted neoantigen-HLA capture reagents. Sixteen patients with different refractory solid cancers received up to three distinct neoTCR transgenic cell products. Each product expressed a patient-specific neoTCR and was administered in a cell-dose-escalation, first-in-human phase I clinical trial ( NCT03970382 ). One patient had grade 1 cytokine release syndrome and one patient had grade 3 encephalitis. All participants had the expected side effects from the lymphodepleting chemotherapy. Five patients had stable disease and the other eleven had disease progression as the best response on the therapy. neoTCR transgenic T cells were detected in tumour biopsy samples after infusion at frequencies higher than the native TCRs before infusion. This study demonstrates the feasibility of isolating and cloning multiple TCRs that recognize mutational neoantigens. Moreover, simultaneous knockout of the endogenous TCR and knock-in of neoTCRs using single-step, non-viral precision genome-editing are achieved. The manufacture of neoTCR engineered T cells at clinical grade, the safety of infusing up to three gene-edited neoTCR T cell products and the ability of the transgenic T cells to traffic to the tumours of patients are also demonstrated.

Financial incentives for vaccination do not have negative unintended consequences.

Financial incentives to encourage healthy and prosocial behaviours often trigger initial behavioural change1-11, but a large academic literature warns against using them12-16. Critics warn that financial incentives can crowd out prosocial motivations and reduce perceived safety and trust, thereby reducing healthy behaviours when no payments are offered and eroding morals more generally17-24. Here we report findings from a large-scale, pre-registered study in Sweden that causally measures the unintended consequences of offering financial incentives for taking the first dose of a COVID-19 vaccine. We use a unique combination of random exposure to financial incentives, population-wide administrative vaccination records and rich survey data. We find no negative consequences of financial incentives; we can reject even small negative impacts of offering financial incentives on future vaccination uptake, morals, trust and perceived safety. In a complementary study, we find that informing US residents about the existence of state incentive programmes also has no negative consequences. Our findings inform not only the academic debate on financial incentives for behaviour change but also policy-makers who consider using financial incentives to change behaviour.

Evaluating eligibility criteria of oncology trials using real-world data and AI.

There is a growing focus on making clinical trials more inclusive but the design of trial eligibility criteria remains challenging1-3. Here we systematically evaluate the effect of different eligibility criteria on cancer trial populations and outcomes with real-world data using the computational framework of Trial Pathfinder. We apply Trial Pathfinder to emulate completed trials of advanced non-small-cell lung cancer using data from a nationwide database of electronic health records comprising 61,094 patients with advanced non-small-cell lung cancer. Our analyses reveal that many common criteria, including exclusions based on several laboratory values, had a minimal effect on the trial hazard ratios. When we used a data-driven approach to broaden restrictive criteria, the pool of eligible patients more than doubled on average and the hazard ratio of the overall survival decreased by an average of 0.05. This suggests that many patients who were not eligible under the original trial criteria could potentially benefit from the treatments. We further support our findings through analyses of other types of cancer and patient-safety data from diverse clinical trials. Our data-driven methodology for evaluating eligibility criteria can facilitate the design of more-inclusive trials while maintaining safeguards for patient safety.

A non-hallucinogenic psychedelic analogue with therapeutic potential.

The psychedelic alkaloid ibogaine has anti-addictive properties in both humans and animals1. Unlike most medications for the treatment of substance use disorders, anecdotal reports suggest that ibogaine has the potential to treat addiction to various substances, including opiates, alcohol and psychostimulants. The effects of ibogaine-like those of other psychedelic compounds-are long-lasting2, which has been attributed to its ability to modify addiction-related neural circuitry through the activation of neurotrophic factor signalling3,4. However, several safety concerns have hindered the clinical development of ibogaine, including its toxicity, hallucinogenic potential and tendency to induce cardiac arrhythmias. Here we apply the principles of function-oriented synthesis to identify the key structural elements of the potential therapeutic pharmacophore of ibogaine, and we use this information to engineer tabernanthalog-a water-soluble, non-hallucinogenic, non-toxic analogue of ibogaine that can be prepared in a single step. In rodents, tabernanthalog was found to promote structural neural plasticity, reduce alcohol- and heroin-seeking behaviour, and produce antidepressant-like effects. This work demonstrates that, through careful chemical design, it is possible to modify a psychedelic compound to produce a safer, non-hallucinogenic variant that has therapeutic potential.

Transplantation Outcomes with Donor Hearts after Circulatory Death.

BACKGROUND: Data showing the efficacy and safety of the transplantation of hearts obtained from donors after circulatory death as compared with hearts obtained from donors after brain death are limited. METHODS: We conducted a randomized, noninferiority trial in which adult candidates for heart transplantation were assigned in a 3:1 ratio to receive a heart after the circulatory death of the donor or a heart from a donor after brain death if that heart was available first (circulatory-death group) or to receive only a heart that had been preserved with the use of traditional cold storage after the brain death of the donor (brain-death group). The primary end point was the risk-adjusted survival at 6 months in the as-treated circulatory-death group as compared with the brain-death group. The primary safety end point was serious adverse events associated with the heart graft at 30 days after transplantation. RESULTS: A total of 180 patients underwent transplantation; 90 (assigned to the circulatory-death group) received a heart donated after circulatory death and 90 (regardless of group assignment) received a heart donated after brain death. A total of 166 transplant recipients were included in the as-treated primary analysis (80 who received a heart from a circulatory-death donor and 86 who received a heart from a brain-death donor). The risk-adjusted 6-month survival in the as-treated population was 94% (95% confidence interval [CI], 88 to 99) among recipients of a heart from a circulatory-death donor, as compared with 90% (95% CI, 84 to 97) among recipients of a heart from a brain-death donor (least-squares mean difference, -3 percentage points; 90% CI, -10 to 3; P<0.001 for noninferiority [margin, 20 percentage points]). There were no substantial between-group differences in the mean per-patient number of serious adverse events associated with the heart graft at 30 days after transplantation. CONCLUSIONS: In this trial, risk-adjusted survival at 6 months after transplantation with a donor heart that had been reanimated and assessed with the use of extracorporeal nonischemic perfusion after circulatory death was not inferior to that after standard-care transplantation with a donor heart that had been preserved with the use of cold storage after brain death. (Funded by TransMedics; ClinicalTrials.gov number, NCT03831048.).

The Safety of Inpatient Health Care.

BACKGROUND: Adverse events during hospitalization are a major cause of patient harm, as documented in the 1991 Harvard Medical Practice Study. Patient safety has changed substantially in the decades since that study was conducted, and a more current assessment of harm during hospitalization is warranted. METHODS: We conducted a retrospective cohort study to assess the frequency, preventability, and severity of patient harm in a random sample of admissions from 11 Massachusetts hospitals during the 2018 calendar year. The occurrence of adverse events was assessed with the use of a trigger method (identification of information in a medical record that was previously shown to be associated with adverse events) and from review of medical records. Trained nurses reviewed records and identified admissions with possible adverse events that were then adjudicated by physicians, who confirmed the presence and characteristics of the adverse events. RESULTS: In a random sample of 2809 admissions, we identified at least one adverse event in 23.6%. Among 978 adverse events, 222 (22.7%) were judged to be preventable and 316 (32.3%) had a severity level of serious (i.e., caused harm that resulted in substantial intervention or prolonged recovery) or higher. A preventable adverse event occurred in 191 (6.8%) of all admissions, and a preventable adverse event with a severity level of serious or higher occurred in 29 (1.0%). There were seven deaths, one of which was deemed to be preventable. Adverse drug events were the most common adverse events (accounting for 39.0% of all events), followed by surgical or other procedural events (30.4%), patient-care events (which were defined as events associated with nursing care, including falls and pressure ulcers) (15.0%), and health care-associated infections (11.9%). CONCLUSIONS: Adverse events were identified in nearly one in four admissions, and approximately one fourth of the events were preventable. These findings underscore the importance of patient safety and the need for continuing improvement. (Funded by the Controlled Risk Insurance Company and the Risk Management Foundation of the Harvard Medical Institutions.).

Illuminating the dark spaces of healthcare with ambient intelligence.

Advances in machine learning and contactless sensors have given rise to ambient intelligence-physical spaces that are sensitive and responsive to the presence of humans. Here we review how this technology could improve our understanding of the metaphorically dark, unobserved spaces of healthcare. In hospital spaces, early applications could soon enable more efficient clinical workflows and improved patient safety in intensive care units and operating rooms. In daily living spaces, ambient intelligence could prolong the independence of older individuals and improve the management of individuals with a chronic disease by understanding everyday behaviour. Similar to other technologies, transformation into clinical applications at scale must overcome challenges such as rigorous clinical validation, appropriate data privacy and model transparency. Thoughtful use of this technology would enable us to understand the complex interplay between the physical environment and health-critical human behaviours.

The promise and challenge of therapeutic genome editing.

Genome editing, which involves the precise manipulation of cellular DNA sequences to alter cell fates and organism traits, has the potential to both improve our understanding of human genetics and cure genetic disease. Here I discuss the scientific, technical and ethical aspects of using CRISPR (clustered regularly interspaced short palindromic repeats) technology for therapeutic applications in humans, focusing on specific examples that highlight both opportunities and challenges. Genome editing is-or will soon be-in the clinic for several diseases, with more applications under development. The rapid pace of the field demands active efforts to ensure that this breakthrough technology is used responsibly to treat, cure and prevent genetic disease.

The Safety of Outpatient Health Care : Review of Electronic Health Records.

BACKGROUND: Despite considerable emphasis on delivering safe care, substantial patient harm occurs. Although most care occurs in the outpatient setting, knowledge of outpatient adverse events (AEs) remains limited. OBJECTIVE: To measure AEs in the outpatient setting. DESIGN: Retrospective review of the electronic health record (EHR). SETTING: 11 outpatient sites in Massachusetts in 2018. PATIENTS: 3103 patients who received outpatient care. MEASUREMENTS: Using a trigger method, nurse reviewers identified possible AEs and physicians adjudicated them, ranked severity, and assessed preventability. Generalized estimating equations were used to assess the association of having at least 1 AE with age, sex, race, and primary insurance. Variation in AE rates was analyzed across sites. RESULTS: The 3103 patients (mean age, 52 years) were more often female (59.8%), White (75.1%), English speakers (90.8%), and privately insured (70.4%) and had a mean of 4 outpatient encounters in 2018. Overall, 7.0% (95% CI, 4.6% to 9.3%) of patients had at least 1 AE (8.6 events per 100 patients annually). Adverse drug events were the most common AE (63.8%), followed by health care-associated infections (14.8%) and surgical or procedural events (14.2%). Severity was serious in 17.4% of AEs, life-threatening in 2.1%, and never fatal. Overall, 23.2% of AEs were preventable. Having at least 1 AE was less often associated with ages 18 to 44 years than with ages 65 to 84 years (standardized risk difference, -0.05 [CI, -0.09 to -0.02]) and more often associated with Black race than with Asian race (standardized risk difference, 0.09 [CI, 0.01 to 0.17]). Across study sites, 1.8% to 23.6% of patients had at least 1 AE and clinical category of AEs varied substantially. LIMITATION: Retrospective EHR review may miss AEs. CONCLUSION: Outpatient harm was relatively common and often serious. Adverse drug events were most frequent. Rates were higher among older adults. Interventions to curtail outpatient harm are urgently needed. PRIMARY FUNDING SOURCE: Controlled Risk Insurance Company and the Risk Management Foundation of the Harvard Medical Institutions.

A streamlined pathway for transcatheter aortic valve implantation: the BENCHMARK study.

BACKGROUND AND AIMS: There is significant potential to streamline the clinical pathway for patients undergoing transcatheter aortic valve implantation (TAVI). The purpose of this study was to evaluate the effect of implementing BENCHMARK best practices on the efficiency and safety of TAVI in 28 sites in 7 European countries. METHODS: This was a study of patients with severe symptomatic aortic stenosis (AS) undergoing TAVI with balloon-expandable valves before and after implementation of BENCHMARK best practices. Principal objectives were to reduce hospital length of stay (LoS) and duration of intensive care stay. Secondary objective was to document patient safety. RESULTS: Between January 2020 and March 2023, 897 patients were documented prior to and 1491 patients after the implementation of BENCHMARK practices. Patient characteristics were consistent with a known older TAVI population and only minor differences. Mean LoS was reduced from 7.7 ± 7.0 to 5.8 ± 5.6 days (median 6 vs. 4 days; P < .001). Duration of intensive care was reduced from 1.8 to 1.3 days (median 1.1 vs. 0.9 days; P < .001). Adoption of peri-procedure best practices led to increased use of local anaesthesia (96.1% vs. 84.3%; P < .001) and decreased procedure (median 47 vs. 60 min; P < .001) and intervention times (85 vs. 95 min; P < .001). Thirty-day patient safety did not appear to be compromised with no differences in all-cause mortality (0.6% in both groups combined), stroke/transient ischaemic attack (1.4%), life-threatening bleeding (1.3%), stage 2/3 acute kidney injury (0.7%), and valve-related readmission (1.2%). CONCLUSIONS: Broad implementation of BENCHMARK practices contributes to improving efficiency of TAVI pathway reducing LoS and costs without compromising patient safety.

Funding and Resourcing Quality Improvement Effort in Infectious Diseases.

Infectious diseases physicians are frequently called on to perform quality improvement and patient safety (QIPS) work. We describe a newly created faculty position at our institution that allows a faculty member with graduate training in quality and safety methodologies to address QIPS priorities at both the division and hospital levels.

Development of Patient Safety Measures to Identify Inappropriate Diagnosis of Common Infections.

BACKGROUND: Inappropriate diagnosis of infections results in antibiotic overuse and may delay diagnosis of underlying conditions. Here we describe the development and characteristics of 2 safety measures of inappropriate diagnosis of urinary tract infection (UTI) and community-acquired pneumonia (CAP), the most common inpatient infections on general medicine services. METHODS: Measures were developed from guidelines and literature and adapted based on data from patients hospitalized with UTI and CAP in 49 Michigan hospitals and feedback from end-users, a technical expert panel (TEP), and a patient focus group. Each measure was assessed for reliability, validity, feasibility, and usability. RESULTS: Two measures, now endorsed by the National Quality Forum (NQF), were developed. Measure reliability (derived from 24 483 patients) was excellent (0.90 for UTI; 0.91 for CAP). Both measures had strong validity demonstrated through (a) face validity by hospital users, the TEPs, and patient focus group, (b) implicit case review (ĸ 0.72 for UTI; ĸ 0.72 for CAP), and (c) rare case misclassification (4% for UTI; 0% for CAP) due to data errors (<2% for UTI; 6.3% for CAP). Measure implementation through hospital peer comparison in Michigan hospitals (2017 to 2020) demonstrated significant decreases in inappropriate diagnosis of UTI and CAP (37% and 32%, respectively, P < .001), supporting usability. CONCLUSIONS: We developed highly reliable, valid, and usable measures of inappropriate diagnosis of UTI and CAP for hospitalized patients. Hospitals seeking to improve diagnostic safety, antibiotic use, and patient care should consider using these measures to reduce inappropriate diagnosis of CAP and UTI.

Pragmatic Approach to In Situ Simulation to Identify Latent Safety Threats Before Moving to a Newly Built ICU.

OBJECTIVES: Transitions to new care environments may have unexpected consequences that threaten patient safety. We undertook a quality improvement project using in situ simulation to learn the new patient care environment and expose latent safety threats before transitioning patients to a newly built adult ICU. DESIGN: Descriptive review of a patient safety initiative. SETTING: A newly built 24-bed neurocritical care unit at a tertiary care academic medical center. SUBJECTS: Care providers working in neurocritical care unit. INTERVENTIONS: We implemented a pragmatic three-stage in situ simulation program to learn a new patient care environment, transitioning patients from an open bay unit to a newly built private room-based ICU. The project tested the safety and efficiency of new workflows created by new patient- and family-centric features of the unit. We used standardized patients and high-fidelity mannequins to simulate patient scenarios, with "test" patients created through all electronic databases. Relevant personnel from clinical and nonclinical services participated in simulations and/or observed scenarios. We held a debriefing after each stage and scenario to identify safety threats and other concerns. Additional feedback was obtained via a written survey sent to all participants. We prospectively surveyed for missed latent safety threats for 2 years following the simulation and fixed issues as they arose. MEASUREMENTS AND MAIN RESULTS: We identified and addressed 70 latent safety threats, including issues concerning physical environment, infection prevention, patient workflow, and informatics before the move into the new unit. We also developed an orientation manual that highlighted new physical and functional features of the ICU and best practices gleaned from the simulations. All participants agreed or strongly agreed that simulations were beneficial. Two-year follow-up revealed only two missed latent safety threats. CONCLUSIONS: In situ simulation effectively identifies latent safety threats surrounding the transition to new ICUs and should be considered before moving into new units.

CRISPR-Cas attack of HIV-1 proviral DNA can cause unintended deletion of surrounding cellular DNA.

IMPORTANCE: Although HIV replication can be effectively inhibited by antiretroviral therapy, this does not result in a cure as the available drugs do not inactivate the integrated HIV-1 DNA in infected cells. Consequently, HIV-infected individuals need lifelong therapy to prevent viral rebound. Several preclinical studies indicate that CRISPR-Cas gene-editing systems can be used to achieve permanent inactivation of the viral DNA. It was previously shown that this inactivation was due to small inactivating mutations at the targeted sites in the HIV genome and to excision or inversion of the viral DNA fragment between two target sites. We, here, demonstrate that CRISPR-Cas treatment also causes large unintended deletions, which can include surrounding chromosomal sequences. As the loss of chromosomal sequences may cause oncogenic transformation of the cell, such unintended large deletions form a potential safety risk in clinical application of this antiviral application and possibly all CRISPR-Cas gene-editing approaches.

What's the CATCH? A Participatory Learning Model for Case Review Rounds to Support a Culture of Safety and Quality Improvement in Pediatric Hospital Medicine.

OBJECTIVE: To evaluate the effectiveness of a new model, Case Analysis and Translation to Care in Hospital (CATCH), for the review of pediatric inpatient cases when an adverse event or "close call" had occurred. STUDY DESIGN: The curricular intervention consisted of an introductory podcast/workshop, mentorship of presenters, and monthly CATCH rounds over 16 months. The study was conducted with 22 pediatricians at a single tertiary care center. Intervention assessment occurred using participant surveys at multiple intervals: pre/post the intervention, presenter experience (post), physicians involved and mentors experience (post), and after each CATCH session. Paired t-tests and thematic analysis were used to analyze data. Time required to support the CATCH process was used to assess feasibility. RESULTS: Our overall experience and data revealed a strong preference for the CATCH model, high levels of engagement and satisfaction with CATCH sessions, and positive presenter as well as physicians-involved and mentor experiences. Participants reported that the CATCH model is feasible, engages physicians, promotes a safe learning environment, facilitates awareness of tools for case analysis, and provides opportunities to create "CATCH of the Day" recommendations to support translation of learning to clinical practice. CONCLUSIONS: The CATCH model has significant potential to strengthen clinical case rounds in pediatric hospital medicine. Future research is needed to assess the effectiveness of the model at additional sites and across medical specialities.

US state laws on medical freedom and investigational stem cell procedures: a call to focus on state-based legislation.

The premature marketing of investigational stem cell interventions (SCIs) is a growing market in the US. Several US states have passed legislation to permit and promote unproven and experimental SCIs for individuals with terminal or chronic diseases. These SCI medical freedom laws, which are largely based on right-to-try legislation, increase access to experimental SCIs with little to no oversight. They undermine federal regulatory authority and can compromise patient safety and informed decision-making. SCI medical freedom laws have gone largely unnoticed by scientific societies interested in the responsible translation of stem cell medicine. In this article, we analyze state SCI medical freedom laws and describe their detrimental impact on patients and society. We contend that scientific and medical societies are uniquely poised to advocate against state-based policy promoting unproven SCIs but recognize resource and other constraints to advocate for or against legislation in 50 states. We recommend societies establish coalitions and share resources to address state-based SCI medical freedom laws and other legislation surrounding unproven SCIs.

"Good Care Is Slow Enough to Be Able to Pay Attention": Primary Care Time Scarcity and Patient Safety.

BACKGROUND: There is growing, widespread recognition that expectations of US primary care vastly exceed the time and resources allocated to it. Little research has directly examined how time scarcity contributes to harm or patient safety incidents not readily capturable by population-based quality metrics. OBJECTIVE: To examine near-miss events identified by primary care physicians in which taking additional time improved patient care or prevented harm. DESIGN: Qualitative study based on semi-structured interviews. PARTICIPANTS: Twenty-five primary care physicians practicing in the USA. APPROACH: Participants completed a survey that included demographic questions, the Ballard Organizational Temporality Scale and the Mini-Z scale, followed by a one hour qualitative interview over video-conference (Zoom). Iterative thematic qualitative data analysis was conducted. KEY RESULTS: Primary care physicians identified several types of near-miss events in which taking extra time during visits changed their clinical management. These were evident in five types of patient care episodes: high-risk social situations, high-risk medication regimens requiring patient education, high acuity conditions requiring immediate workup or treatment, interactions of physical and mental health, and investigating more subtle clinical suspicions. These near-miss events highlight the ways in which unreasonably large patient panels and packed schedules impede adequate responses to patient care episodes that are time sensitive and intensive or require flexibility. CONCLUSIONS: Primary care physicians identify and address patient safety issues and high-risk situations by spending more time than allotted for a given patient encounter. Current quality metrics do not account for this critical aspect of primary care work. Current healthcare policy and organization create time scarcity. Interventions to address time scarcity and to measure its prevalence and implications for care quality and safety are urgently needed.

Device regulation and surveillance in vascular care: Challenges and opportunities.

Cardiovascular devices are essential for the treatment of cardiovascular diseases including cerebrovascular, coronary, valvular, congenital, peripheral vascular and arrhythmic diseases. The regulation and surveillance of vascular devices in real-world practice, however, presents challenges during each individual product's life cycle. Four examples illustrate recent challenges and questions regarding safety, appropriate use and efficacy arising from FDA approved devices used in real-world practice. We outline potential pathways wherein providers, regulators and payors could potentially provide high-quality cardiovascular care, identify safety signals, ensure equitable device access, and study potential issues with devices in real-world practice.

Patient safety incidents in endoscopy: a human factors analysis of nonprocedural significant harm incidents from the National Reporting and Learning System (NRLS).

BACKGROUND: Despite advances in understanding and reducing the risk of endoscopic procedures, there is little consideration of the safety of the wider endoscopy service. Patient safety incidents (PSIs) still occur. We sought to identify nonprocedural PSIs (nPSIs) and their causative factors from a human factors perspective and generate ideas for safety improvement. METHODS: Endoscopy-specific PSI reports were extracted from the National Reporting and Learning System (NRLS). A retrospective, cross-sectional human factors analysis of data was performed. Two independent researchers coded data using a hybrid thematic analysis approach. The Human Factors Analysis and Classification System (HFACS) was used to code contributory factors. Analysis informed creation of driver diagrams and key recommendations for safety improvement in endoscopy. RESULTS: From 2017 to 2019, 1181 endoscopy-specific PSIs of significant harm were reported across England and Wales, with 539 (45.6%) being nPSIs. Five categories accounted for over 80% of all incidents, with "follow-up and surveillance" being the largest (23.4% of all nPSIs). From the free-text incident reports, 487 human factors codes were identified. Decision-based errors were the most common act prior to PSI occurrence. Other frequent preconditions to incidents were focused on environmental factors, particularly overwhelmed resources, patient factors, and ineffective team communication. Lack of staffing, standard operating procedures, effective systems, and clinical pathways were also contributory. Seven key recommendations for improving safety have been made in response to our findings. CONCLUSIONS: This was the first national-level human factors analysis of endoscopy-specific PSIs. This work will inform safety improvement strategies and should empower individual services to review their approach to safety.