Research has found that artificial intelligence (AI) can improve diagnostic accuracy, but less is known about its performance in clinical settings. To evaluate the performance of AI in identifying breast cancer in a clinical setting, this study deployed AI in a curated, representative data set from the UK (25,856 women) and an enriched dataset from the US (3,097 women), as well as compared the performance of AI to that of six human radiologist readers. They used biopsy-confirmed cancer patients to evaluate AI predictions. The authors reported a reduction in both false positives and false negatives using AI and found that the AI system was more accurate than the radiologists.

Effective teamwork is a critical component of care coordination and patient safety. This systematic review assessed the relationship between teamwork processes and clinical and process outcome measures in an acute care setting. Outcome measures included both clinical outcomes, such as postoperative infection rates, and process measures, such as adherence to checklists intended to prevent patient harm. The authors found that teamwork was positively correlated with both outcome and process measures, regardless of the characteristics of the team or task.

Burnout has been linked to medical errors in a variety of clinician environments. This commentary describes burnout as an occupational problem in health care and recommends assessment at the departmental level to achieve lasting change.

The extent of harm due to patient safety problems varies across studies. This systematic review sought to estimate the prevalence of preventable harm in medical care overall. Researchers synthesized data from 70 studies and estimated that 6% of patients receiving medical care experience preventable harm. Harm related to medications, diagnosis, health care–associated infections, and procedures accounted for significant proportions of preventable harm. The authors conclude that focusing on evidenced-based strategies to address preventable patient harm would improve health care quality and subsequently reduce costs. A related editorial calls for improving measurement of preventable harm. Another editorial spotlights the importance of understanding the causes of preventable harm in health care.

Machine learning may have the potential to improve clinical decision-making and diagnosis. In this study, machine-learning algorithms generally performed better than human experts in accurately diagnosing 7 types of pigmented skin lesions and the top 3 algorithms performed better than the 27 physicians.

High-risk opioid prescribing by providers contributes to opioid misuse. This systematic review sought to identify factors that confer risk for opioid addiction and thereby suggest which patients can safely take opioids. Researchers found that a prior history of substance use disorder, prescription of psychiatric medications, certain mental health diagnoses, higher daily opioid doses, and prescription of opioids for 30 days or more may confer risk for opioid addiction. The only factor associated with a lower risk of opioid use disorder was absence of a mood disorder. They could not identify any screening instruments or tools that accurately risk-stratified individuals' likelihood of opioid addiction. An Annual Perspective discussed problematic prescribing practices that likely contribute to adverse events and described promising practices to foster safer opioid use.

Checklists have been shown to improve surgical safety in randomized controlled trials, but they have had varied impact when implemented in clinical practice. This interrupted time-series study examined surgical mortality before, during, and after implementation of the WHO surgical safety checklist. The rate of surgical mortality declined more during checklist introduction than it had before or after implementation, and hospital mortality did not decline among nonsurgical patients during the same time interval. The investigators, including checklist pioneer Atul Gawande, conclude that perioperative mortality has declined in association with checklist implementation. Past PSNet interviews with Lucian Leape and David Urbach discussed their perspectives on surgical safety checklists.

Artificial intelligence (AI) may have the capacity to improve diagnosis. Researchers found that an AI system was able to detect breast cancer using mammography with accuracy similar to that of the average of the 101 radiologists whose interpretations were included in the study.

Hospital-acquired pressure ulcers are considered never events and represent a significant source of patient harm. This systematic review and meta-analysis found a pooled incidence of 12% for pressure injuries related to the use of medical devices.

Clinician burnout has been associated with decreased job satisfaction. Burnout may also be detrimental to patient safety. This systematic review and meta-analysis found high rates of burnout among mental health professionals. The authors recommend strategies to address burnout including promoting professional autonomy, developing teamwork, and providing quality clinical supervision.

Standardized handoff tools are increasingly implemented to improve communication between health care providers. Although this systematic review identified several studies supporting the use of SBAR as a communication tool to improve patient safety, the authors suggest the evidence is moderate and that further research is needed.

Cognitive biases can lead to unnecessary treatment and delays in diagnosis. This commentary reviews examples of bias that commonly occur in medical practice and describes debiasing tactics to help improve decision-making.

Inadequate hospital nurse staffing is linked to increased mortality. This systematic review found that lower nurse staffing is associated with more reports of missed nursing care. Two of the authors summarized the science of missed nursing care in a recent PSNet perspective.

Measuring patient safety is critical to improvement. This ethnographic study examined the implementation of a patient safety measurement program in the United Kingdom, the NHS Safety Thermometer, which measured incidence of pressure ulcers, harm from falls, catheter-associated urinary tract infection, and venous thromboembolism, with the goal of informing local improvement efforts. Investigators sought to examine how the measurement program was perceived by frontline staff. Despite the explicit emphasis on using the data for improvement, it was viewed as an external reporting requirement. The program was also viewed as a basis to compare organizations, especially because it included pay-for-performance incentives. The authors suggest that the intention of the program did not match the real-world considerations of participating health care systems and had the unintended consequence of creating potential for blame.

Identifying incipient clinical deterioration is a prerequisite for rapid response and prevention of harm for hospitalized patients. This study tested a bedside pediatric early warning system, which included an illness severity score, standardized documentation, and monitoring protocols. In a cluster-randomized trial in several high-income countries, implementation of the bundle did not result in decreased in-hospital mortality compared to usual care. The overall mortality rate in the study was less than 0.2%. The authors suggest that this unexpectedly low mortality rate may have made it difficult to detect differences in intervention versus control hospitals. A related editorial suggests that artificial intelligence should be used to identify clinical deterioration and that outcomes beyond mortality should be considered in their evaluation.

Preventable harm is common during and after hospital discharge. Pharmacist-delivered medication reconciliation has been proposed as a strategy to reduce adverse medication events and readmissions. Investigators conducted a three-arm randomized controlled trial comparing the effect of pharmacist-delivered medication reviews, motivational interviews, and postdischarge follow-up with nursing homes, primary care providers, and pharmacies (extended intervention); simple inpatient medication reconciliation (basic intervention); and usual care (no intervention) on outcomes for medically complex patients. The extended intervention reduced hospital readmissions and emergency department visits within 180 days of discharge while the basic intervention did not. This trial was large, robustly conducted, and demonstrated a durable improvement in safety for patients at increased readmission risk. A previous Annual Perspective explored tools for safer transitions of care.

The field of patient safety has long looked to high reliability organizations like aviation or nuclear power for solutions, but it is unclear how well such approaches translate to health care settings. In this study, researchers asked clinicians to identify safety hazards from their own work and then propose solutions. After applying a systems thinking framework to clinicians' solution ideas, they found that most of the clinician-generated safety approaches would be considered ineffective by high reliability standards. The authors suggest that industrial frameworks are an imperfect match for health care settings and should be used with caution. A recent PSNet interview with the study's senior author, Mary Dixon-Woods, discusses the sociology of health care versus other industries.

Diagnostic error is a growing area of focus within patient safety. Artificial intelligence has the potential to improve the diagnostic process, both in terms of accuracy and efficiency. In this study, investigators compared the use of automated deep learning algorithms for detecting metastatic disease in stained tissue sections of lymph nodes of women with breast cancer to pathologists' diagnoses. The algorithms were developed by researchers as part of a competition and their performance was assessed on a test set of 129 slides, 49 with metastatic disease and 80 without. A panel of 11 pathologists evaluated the same slides with a 2-hour time limit and one pathologist evaluated the slides without any time constraints. The authors conclude that some of the algorithms demonstrated better diagnostic performance than the pathologists did, but they suggest that further testing in a clinical setting is warranted. An accompanying editorial discusses the potential of artificial intelligence in health care.

In-hospital mortality for many conditions is higher on the weekends than on weekdays—a phenomenon known as the weekend effect. Some hypothesize lower specialty physician staffing levels on weekends explains the mortality difference. This cross-sectional study compared specialist staffing levels and mortality rates at 115 hospitals in the English National Health Service on Sundays compared to Wednesdays. Researchers found a higher mortality rate and lower intensity of specialty services on weekends, but there was no correlation between the two ratios. Although this study is not definitive, it does imply that alternate mechanisms may explain the weekend effect, such as case mix differences, variation in nonphysician staffing, or lower availability of diagnostic services. A previous PSNet interview discussed the weekend effect in health care.

Missed nursing care may result from inadequate nurse staffing and explain the relationship between nurse-to-patient ratios and patient outcomes. Research has shown that higher nurse staffing levels are associated with lower inpatient mortality and that reduced staffing increases the risk for postoperative complications. In this study, investigators examined data from more than 400,000 surgical patients from 300 hospitals in 9 countries as well as survey responses from 26,516 nurses. They found a significant association between nurse staffing and missed nursing care with 30-day risk-adjusted postoperative mortality. The authors conclude that measuring missed nursing care may help identify patients at greater risk for adverse outcomes earlier in their course. A past WebM&M commentary highlighted important issues associated with nurse staffing ratios.