Building a house without foundations? A 24-country qualitative interview study on artificial intelligence in intensive care medicine

OBJECTIVES

To explore the views of intensive care professionals in high-income countries (HICs) and lower-to-middle-income countries (LMICs) regarding the use and implementation of artificial intelligence (AI) technologies in intensive care units (ICUs).

METHODS

Individual semi-structured qualitative interviews were conducted between December 2021 and August 2022 with 59 intensive care professionals from 24 countries. Transcripts were analysed using conventional content analysis.

RESULTS

Participants had generally positive views about the potential use of AI in ICUs but also reported some well-known concerns about the use of AI in clinical practice and important technical and non-technical barriers to the implementation of AI. Important differences existed between ICUs regarding their current readiness to implement AI. However, these differences were not primarily between HICs and LMICs, but between a small number of ICUs in large tertiary hospitals in HICs, which were reported to have the necessary digital infrastructure for AI, and nearly all other ICUs in both HICs and LMICs, which were reported to neither have the technical capability to capture the necessary data or use AI, nor the staff with the right knowledge and skills to use the technology.

CONCLUSION

Pouring massive amounts of resources into developing AI without first building the necessary digital infrastructure foundation needed for AI is unethical. Real-world implementation and routine use of AI in the vast majority of ICUs in both HICs and LMICs included in our study is unlikely to occur any time soon. ICUs should not be using AI until certain preconditions are met.

Variation in monitoring: Glucose measurement in the ICU as a case study to preempt spurious correlations

OBJECTIVE

Health inequities can be influenced by demographic factors such as race and ethnicity, proficiency in English, and biological sex. Disparities may manifest as differential likelihood of testing which correlates directly with the likelihood of an intervention to address an abnormal finding. Our retrospective observational study evaluated the presence of variation in glucose measurements in the Intensive Care Unit (ICU).

METHODS

Using the MIMIC-IV database (2008-2019), a single-center, academic referral hospital in Boston (USA), we identified adult patients meeting sepsis-3 criteria. Exclusion criteria were diabetic ketoacidosis, ICU length of stay under 1 day, and unknown race or ethnicity. We performed a logistic regression analysis to assess differential likelihoods of glucose measurements on day 1. A negative binomial regression was fitted to assess the frequency of subsequent glucose readings. Analyses were adjusted for relevant clinical confounders, and performed across three disparity proxy axes: race and ethnicity, sex, and English proficiency.

RESULTS

We studied 24,927 patients, of which 19.5% represented racial and ethnic minority groups, 42.4% were female, and 9.8% had limited English proficiency. No significant differences were found for glucose measurement on day 1 in the ICU. This pattern was consistent irrespective of the axis of analysis, i.e. race and ethnicity, sex, or English proficiency. Conversely, subsequent measurement frequency revealed potential disparities. Specifically, males (incidence rate ratio (IRR) 1.06, 95% confidence interval (CI) 1.01 - 1.21), patients who identify themselves as Hispanic (IRR 1.11, 95% CI 1.01 - 1.21), or Black (IRR 1.06, 95% CI 1.01 - 1.12), and patients being English proficient (IRR 1.08, 95% CI 1.01 - 1.15) had higher chances of subsequent glucose readings.

CONCLUSION

We found disparities in ICU glucose measurements among patients with sepsis, albeit the magnitude was small. Variation in disease monitoring is a source of data bias that may lead to spurious correlations when modeling health data.

Can large language models provide secondary reliable opinion on treatment options for dermatological diseases?

OBJECTIVE

To investigate the consistency and reliability of medication recommendations provided by ChatGPT for common dermatological conditions, highlighting the potential for ChatGPT to offer second opinions in patient treatment while also delineating possible limitations.

MATERIALS AND METHODS

In this mixed-methods study, we used survey questions in April 2023 for drug recommendations generated by ChatGPT with data from secondary databases, that is, Taiwan's National Health Insurance Research Database and an US medical center database, and validated by dermatologists. The methodology included preprocessing queries, executing them multiple times, and evaluating ChatGPT responses against the databases and dermatologists. The ChatGPT-generated responses were analyzed statistically in a disease-drug matrix, considering disease-medication associations (Q-value) and expert evaluation.

RESULTS

ChatGPT achieved a high 98.87% dermatologist approval rate for common dermatological medication recommendations. We evaluated its drug suggestions using the Q-value, showing that human expert validation agreement surpassed Q-value cutoff-based agreement. Varying cutoff values for disease-medication associations, a cutoff of 3 achieved 95.14% accurate prescriptions, 5 yielded 85.42%, and 10 resulted in 72.92%. While ChatGPT offered accurate drug advice, it occasionally included incorrect ATC codes, leading to issues like incorrect drug use and type, nonexistent codes, repeated errors, and incomplete medication codes.

CONCLUSION

ChatGPT provides medication recommendations as a second opinion in dermatology treatment, but its reliability and comprehensiveness need refinement for greater accuracy. In the future, integrating a medical domain-specific knowledge base for training and ongoing optimization will enhance the precision of ChatGPT's results.

Critical appraisal of machine learning prognostic models for acute pancreatitis: protocol for a systematic review

Acute pancreatitis (AP) is an acute inflammatory disorder that is common, costly, and is increasing in incidence worldwide with over 300,000 hospitalizations occurring yearly in the United States alone. As its course and outcomes vary widely, a critical knowledge gap in the field has been a lack of accurate prognostic tools to forecast AP patients' outcomes. Despite several published studies in the last three decades, the predictive performance of published prognostic models has been found to be suboptimal. Recently, non-regression machine learning models (ML) have garnered intense interest in medicine for their potential for better predictive performance. Each year, an increasing number of AP models are being published. However, their methodologic quality relating to transparent reporting and risk of bias in study design has never been systematically appraised. Therefore, through collaboration between a group of clinicians and data scientists with appropriate content expertise, we will perform a systematic review of papers published between January 2021 and December 2023 containing artificial intelligence prognostic models in AP. To systematically assess these studies, the authors will leverage the CHARMS checklist, PROBAST tool for risk of bias assessment, and the most current version of the TRIPOD-AI. (Research Registry ( http://www.reviewregistry1727 .).

Use of Artificial Intelligence in Improving Outcomes in Heart Disease: A Scientific Statement From the American Heart Association

A major focus of academia, industry, and global governmental agencies is to develop and apply artificial intelligence and other advanced analytical tools to transform health care delivery. The American Heart Association supports the creation of tools and services that would further the science and practice of precision medicine by enabling more precise approaches to cardiovascular and stroke research, prevention, and care of individuals and populations. Nevertheless, several challenges exist, and few artificial intelligence tools have been shown to improve cardiovascular and stroke care sufficiently to be widely adopted. This scientific statement outlines the current state of the art on the use of artificial intelligence algorithms and data science in the diagnosis, classification, and treatment of cardiovascular disease. It also sets out to advance this mission, focusing on how digital tools and, in particular, artificial intelligence may provide clinical and mechanistic insights, address bias in clinical studies, and facilitate education and implementation science to improve cardiovascular and stroke outcomes. Last, a key objective of this scientific statement is to further the field by identifying best practices, gaps, and challenges for interested stakeholders.

Ophthalmology Optical Coherence Tomography Databases for Artificial Intelligence Algorithm: A Review

BACKGROUND

Imaging plays a pivotal role in eye assessment. With the introduction of advanced machine learning and artificial intelligence (AI), the focus has shifted to imaging datasets in ophthalmology. While disparities and health inequalities hidden within data are well-documented, the ophthalmology field faces specific challenges to the creation and maintenance of datasets. Optical Coherence Tomography (OCT) is useful for the diagnosis and monitoring of retinal pathologies, making it valuable for AI applications. This review aims to identify and compare the landscape of publicly available optical coherence tomography databases for AI applications.

METHODS

We conducted a literature review on OCT and AI articles with publicly accessible datasets, using PubMed, Scopus, and Web of Science databases. The review retrieved 183 articles, and after full-text analysis, 50 articles were included. From the included articles were identified 8 publicly available OCT datasets, focusing on patient demographics and clinical details for thorough assessment and comparison.

RESULTS

The resulting datasets encompass 154,313 images collected from Spectralis, Cirrus HD, Topcon 3D, and Bioptigen devices. These datasets included normal exams, age-related macular degeneration, and diabetic maculopathy, among others. Comprehensive demographic information is available in one dataset and the USA is the most represented population.

DISCUSSION

Current publicly available OCT databases for AI applications exhibit limitations, stemming from their non-representative nature and the lack of comprehensive demographic information. Limited datasets hamper research and equitable AI development. To promote equitable AI algorithmic development in ophthalmology, there is a need for the creation and dissemination of more representative datasets.

Towards long-tailed, multi-label disease classification from chest X-ray: Overview of the CXR-LT challenge

Many real-world image recognition problems, such as diagnostic medical imaging exams, are "long-tailed" - there are a few common findings followed by many more relatively rare conditions. In chest radiography, diagnosis is both a long-tailed and multi-label problem, as patients often present with multiple findings simultaneously. While researchers have begun to study the problem of long-tailed learning in medical image recognition, few have studied the interaction of label imbalance and label co-occurrence posed by long-tailed, multi-label disease classification. To engage with the research community on this emerging topic, we conducted an open challenge, CXR-LT, on long-tailed, multi-label thorax disease classification from chest X-rays (CXRs). We publicly release a large-scale benchmark dataset of over 350,000 CXRs, each labeled with at least one of 26 clinical findings following a long-tailed distribution. We synthesize common themes of top-performing solutions, providing practical recommendations for long-tailed, multi-label medical image classification. Finally, we use these insights to propose a path forward involving vision-language foundation models for few- and zero-shot disease classification.

Machine learning for healthcare that matters: Reorienting from technical novelty to equitable impact

Despite significant technical advances in machine learning (ML) over the past several years, the tangible impact of this technology in healthcare has been limited. This is due not only to the particular complexities of healthcare, but also due to structural issues in the machine learning for healthcare (MLHC) community which broadly reward technical novelty over tangible, equitable impact. We structure our work as a healthcare-focused echo of the 2012 paper "Machine Learning that Matters", which highlighted such structural issues in the ML community at large, and offered a series of clearly defined "Impact Challenges" to which the field should orient itself. Drawing on the expertise of a diverse and international group of authors, we engage in a narrative review and examine issues in the research background environment, training processes, evaluation metrics, and deployment protocols which act to limit the real-world applicability of MLHC. Broadly, we seek to distinguish between machine learning ON healthcare data and machine learning FOR healthcare-the former of which sees healthcare as merely a source of interesting technical challenges, and the latter of which regards ML as a tool in service of meeting tangible clinical needs. We offer specific recommendations for a series of stakeholders in the field, from ML researchers and clinicians, to the institutions in which they work, and the governments which regulate their data access.

Participant flow diagrams for health equity in AI

Selection bias can arise through many aspects of a study, including recruitment, inclusion/exclusion criteria, input-level exclusion and outcome-level exclusion, and often reflects the underrepresentation of populations historically disadvantaged in medical research. The effects of selection bias can be further amplified when non-representative samples are used in artificial intelligence (AI) and machine learning (ML) applications to construct clinical algorithms. Building on the "Data Cards" initiative for transparency in AI research, we advocate for the addition of a participant flow diagram for AI studies detailing relevant sociodemographic and/or clinical characteristics of excluded participants across study phases, with the goal of identifying potential algorithmic biases before their clinical implementation. We include both a model for this flow diagram as well as a brief case study explaining how it could be implemented in practice. Through standardized reporting of participant flow diagrams, we aim to better identify potential inequities embedded in AI applications, facilitating more reliable and equitable clinical algorithms.

Drop the shortcuts: image augmentation improves fairness and decreases AI detection of race and other demographics from medical images

BACKGROUND

It has been shown that AI models can learn race on medical images, leading to algorithmic bias. Our aim in this study was to enhance the fairness of medical image models by eliminating bias related to race, age, and sex. We hypothesise models may be learning demographics via shortcut learning and combat this using image augmentation.

METHODS

This study included 44,953 patients who identified as Asian, Black, or White (mean age, 60.68 years ±18.21; 23,499 women) for a total of 194,359 chest X-rays (CXRs) from MIMIC-CXR database. The included CheXpert images comprised 45,095 patients (mean age 63.10 years ±18.14; 20,437 women) for a total of 134,300 CXRs were used for external validation. We also collected 1195 3D brain magnetic resonance imaging (MRI) data from the ADNI database, which included 273 participants with an average age of 76.97 years ±14.22, and 142 females. DL models were trained on either non-augmented or augmented images and assessed using disparity metrics. The features learned by the models were analysed using task transfer experiments and model visualisation techniques.

FINDINGS

In the detection of radiological findings, training a model using augmented CXR images was shown to reduce disparities in error rate among racial groups (-5.45%), age groups (-13.94%), and sex (-22.22%). For AD detection, the model trained with augmented MRI images was shown 53.11% and 31.01% reduction of disparities in error rate among age and sex groups, respectively. Image augmentation led to a reduction in the model's ability to identify demographic attributes and resulted in the model trained for clinical purposes incorporating fewer demographic features.

INTERPRETATION

The model trained using the augmented images was less likely to be influenced by demographic information in detecting image labels. These results demonstrate that the proposed augmentation scheme could enhance the fairness of interpretations by DL models when dealing with data from patients with different demographic backgrounds.

FUNDING

National Science and Technology Council (Taiwan), National Institutes of Health.

Disparities in access to and timing of interventional therapies for pulmonary embolism across the United States

BACKGROUND

Interventional therapies (ITs) are an emerging treatment modality for pulmonary embolism (PE); however, the degree of racial, sex-based, and sociodemographic disparities in access and timing is unknown.

OBJECTIVES

To investigate barriers to access and timing of ITs for PE across the United States.

METHODS

A retrospective cohort study utilizing the Nationwide Inpatient Sample from 2016-2020 included adult patients with PE. The use of ITs (mechanical thrombectomy and catheter-directed thrombolysis) was identified via International Classification of Diseases 10th revision codes. Early IT was defined as procedure performed within the first 2 days after admission.

RESULTS

A total of 27 805 273 records from the 2016-2020 Nationwide Inpatient Sample database were examined. There were 387 514 (1.4%) patients with PE, with 14 249 (3.6%) of them having undergone IT procedures (11 115 catheter-directed thrombolysis, 2314 thrombectomy, and 780 both procedures). After multivariate adjustment, factors associated with less use of IT included Black race (odds ratio [OR], 0.90; 95% CI, 0.86-0.94; P < .01), Hispanic race (OR, 0.73; 95% CI, 0.68-0.79; P < .01), female sex (OR, 0.88; 95% CI, 0.85-0.91; P < .01), treatment in a rural hospital (OR, 0.49; 95% CI, 0.44-0.54; P < .01), and lack of private insurance (Medicare OR, 0.77; 95% CI, 0.73-0.80; P < .01; Medicaid OR, 0.65; 95% CI, 0.61-0.69; P < .01; no coverage OR, 0.87; 95% CI, 0.82-0.93; P < .01). Among the patients who received IT, 11 315 (79%) procedures were conducted within 2 days of admission and 2934 (21%) were delayed. Factors associated with delayed procedures included Black race (OR, 1.12; 95% CI, 1.01-1.26; P = .04), Hispanic race (OR, 1.52; 95% CI, 1.28-1.80; P < .01), weekend admission (OR, 1.37; 95% CI, 1.25-1.51; P < .01), Medicare coverage (OR, 1.24; 95% CI, 1.10-1.40; P < .01), and Medicaid coverage (OR, 1.29; 95% CI, 1.12-1.49; P < .01).

CONCLUSION

Significant racial, sex-based, and geographic barriers exist in overall access to IT for PE in the United States.

Does diversity beget diversity? A scientometric analysis of over 150,000 studies and 49,000 authors published in high-impact medical journals between 2007 and 2022

Background

Health research that significantly impacts global clinical practice and policy is often published in high-impact factor (IF) medical journals. These outlets play a pivotal role in the worldwide dissemination of novel medical knowledge. However, researchers identifying as women and those affiliated with institutions in low- and middle-income countries (LMIC) have been largely underrepresented in high-IF journals across multiple fields of medicine. To evaluate disparities in gender and geographical representation among authors who have published in any of five top general medical journals, we conducted scientometric analyses using a large-scale dataset extracted from the New England Journal of Medicine (NEJM), Journal of the American Medical Association (JAMA), The British Medical Journal (BMJ), The Lancet, and Nature Medicine.

Methods

Author metadata from all articles published in the selected journals between 2007 and 2022 were collected using the DimensionsAI platform. The Genderize.io API was then utilized to infer each author's likely gender based on their extracted first name. The World Bank country classification was used to map countries associated with researcher affiliations to the LMIC or the high-income country (HIC) category. We characterized the overall gender and country income category representation across the medical journals. In addition, we computed article-level diversity metrics and contrasted their distributions across the journals.

Findings

We studied 151,536 authors across 49,764 articles published in five top medical journals, over a long period spanning 15 years. On average, approximately one-third (33.1%) of the authors of a given paper were inferred to be women; this result was consistent across the journals we studied. Further, 86.6% of the teams were exclusively composed of HIC authors; in contrast, only 3.9% were exclusively composed of LMIC authors. The probability of serving as the first or last author was significantly higher if the author was inferred to be a man (18.1% vs 16.8%, P < .01) or was affiliated with an institution in a HIC (16.9% vs 15.5%, P < .01). Our primary finding reveals that having a diverse team promotes further diversity, within the same dimension (i.e., gender or geography) and across dimensions. Notably, papers with at least one woman among the authors were more likely to also involve at least two LMIC authors (11.7% versus 10.4% in baseline, P < .001; based on inferred gender); conversely, papers with at least one LMIC author were more likely to also involve at least two women (49.4% versus 37.6%, P < .001; based on inferred gender).

Conclusion

We provide a scientometric framework to assess authorship diversity. Our research suggests that the inclusiveness of high-impact medical journals is limited in terms of both gender and geography. We advocate for medical journals to adopt policies and practices that promote greater diversity and collaborative research. In addition, our findings offer a first step towards understanding the composition of teams conducting medical research globally and an opportunity for individual authors to reflect on their own collaborative research practices and possibilities to cultivate more diverse partnerships in their work.

A Reassessment of Sodium Correction Rates and Hospital Length of Stay Accounting for Admission Diagnosis

Background

Recent studies have challenged assumptions about slow correction of severe hyponatremia and have shown that rapid correction is associated with shorter hospital length of stay. However, the confounding effect of admission diagnosis has not been fully explored. The objective of this study was to determine whether rapid correction is still associated with shorter length of stay when controlling for admission diagnosis.

Methods

This retrospective cohort study is based on the Medical Information Mart for Intensive Care, including data from both MIMIC-III (2001-2012) and MIMIC-IV (2008-2019). Patients were identified who presented to the hospital with initial sodium <120 mEq/L and were categorized according to total sodium correction achieved in the first day (<6 mEq/L; 6-10 mEq/L; >10 mEq/L). Linear regression was used to assess for an association between correction rate and hospital length of stay, and to determine if this association was significant when controlling for admission diagnosis classifications based on diagnosis related groups (DRGs).

Results

There were 636 patients included in this study. Median [IQR] hospital length of stay was 7 [4, 11] days. Patients had a median [IQR] initial sodium value of 117 [114, 118] mEq/L and final sodium value of 124 [119, 128] mEq/L. In a univariate linear regression, the highest rate of correction (>10 mEq/L) was associated with a shorter length of stay than a moderate rate of correction (coef. -2.363, 95% CI [-4.710, -0.017], p=0.048), but the association was not significant when controlling for admission diagnosis group (coef. -1.685, 95% CI [-3.836, 0.467], p=0.125).

Conclusions

Faster sodium correction was not associated with shorter length of stay when controlling for admission diagnosis categories, suggesting that the disease state confounds this association. While some patients may be discharged earlier if sodium is corrected more rapidly, others may not benefit or may be harmed by this strategy.

Image quality assessment of retinal fundus photographs for diabetic retinopathy in the machine learning era: a review

This study aimed to evaluate the image quality assessment (IQA) and quality criteria employed in publicly available datasets for diabetic retinopathy (DR). A literature search strategy was used to identify relevant datasets, and 20 datasets were included in the analysis. Out of these, 12 datasets mentioned performing IQA, but only eight specified the quality criteria used. The reported quality criteria varied widely across datasets, and accessing the information was often challenging. The findings highlight the importance of IQA for AI model development while emphasizing the need for clear and accessible reporting of IQA information. The study suggests that automated quality assessments can be a valid alternative to manual labeling and emphasizes the importance of establishing quality standards based on population characteristics, clinical use, and research purposes. In conclusion, image quality assessment is important for AI model development; however, strict data quality standards must not limit data sharing. Given the importance of IQA for developing, validating, and implementing deep learning (DL) algorithms, it's recommended that this information be reported in a clear, specific, and accessible way whenever possible. Automated quality assessments are a valid alternative to the traditional manual labeling process, and quality standards should be determined according to population characteristics, clinical use, and research purpose.

Association between Patient Race and Ethnicity and Use of Invasive Ventilation in the United States

Rationale: Outcomes for people with respiratory failure in the United States vary by patient race and ethnicity. Invasive ventilation is an important treatment initiated based on expert opinion. It is unknown whether the use of invasive ventilation varies by patient race and ethnicity. Objectives: To measure 1) the association between patient race and ethnicity and the use of invasive ventilation; and 2) the change in 28-day mortality mediated by any association. Methods: We performed a multicenter cohort study of nonintubated adults receiving oxygen within 24 hours of intensive care admission using the Medical Information Mart for Intensive Care IV (MIMIC-IV, 2008-2019) and Phillips eICU (eICU, 2014-2015) databases from the United States. We modeled the association between patient race and ethnicity (Asian, Black, Hispanic, White) and invasive ventilation rate using a Bayesian multistate model that adjusted for baseline and time-varying covariates, calculated hazard ratios (HRs), and estimated 28-day hospital mortality changes mediated by differential invasive ventilation use. We reported posterior means and 95% credible intervals (CrIs). Results: We studied 38,258 patients, 52% (20,032) from MIMIC-IV and 48% (18,226) from eICU: 2% Asian (892), 11% Black (4,289), 5% Hispanic (1,964), and 81% White (31,113). Invasive ventilation occurred in 9.2% (3,511), and 7.5% (2,869) died. The adjusted rate of invasive ventilation was lower in Asian (HR, 0.82; CrI, 0.70-0.95), Black (HR, 0.78; CrI, 0.71-0.86), and Hispanic (HR, 0.70; CrI, 0.61-0.79) patients compared with White patients. For the average patient, lower rates of invasive ventilation did not mediate differences in 28-day mortality. For a patient on high-flow nasal cannula with inspired oxygen fraction of 1.0, the odds ratios for mortality if invasive ventilation rates were equal to the rate for White patients were 0.97 (CrI, 0.91-1.03) for Asian patients, 0.96 (CrI, 0.91-1.03) for Black patients, and 0.94 (CrI, 0.89-1.01) for Hispanic patients. Conclusions: Asian, Black, and Hispanic patients had lower rates of invasive ventilation than White patients. These decreases did not mediate harm for the average patient, but we could not rule out harm for patients with more severe hypoxemia.

BRSET: A Brazilian Multilabel Ophthalmological Dataset of Retina Fundus Photos

Introduction

The Brazilian Multilabel Ophthalmological Dataset (BRSET) addresses the scarcity of publicly available ophthalmological datasets in Latin America. BRSET comprises 16,266 color fundus retinal photos from 8,524 Brazilian patients, aiming to enhance data representativeness, serving as a research and teaching tool. It contains sociodemographic information, enabling investigations into differential model performance across demographic groups.

Methods

Data from three São Paulo outpatient centers yielded demographic and medical information from electronic records, including nationality, age, sex, clinical history, insulin use, and duration of diabetes diagnosis. A retinal specialist labeled images for anatomical features (optic disc, blood vessels, macula), quality control (focus, illumination, image field, artifacts), and pathologies (e.g., diabetic retinopathy). Diabetic retinopathy was graded using International Clinic Diabetic Retinopathy and Scottish Diabetic Retinopathy Grading. Validation used Dino V2 Base for feature extraction, with 70% training and 30% testing subsets. Support Vector Machines (SVM) and Logistic Regression (LR) were employed with weighted training. Performance metrics included area under the receiver operating curve (AUC) and Macro F1-score.

Results

BRSET comprises 65.1% Canon CR2 and 34.9% Nikon NF5050 images. 61.8% of the patients are female, and the average age is 57.6 years. Diabetic retinopathy affected 15.8% of patients, across a spectrum of disease severity. Anatomically, 20.2% showed abnormal optic discs, 4.9% abnormal blood vessels, and 28.8% abnormal macula. Models were trained on BRSET in three prediction tasks: "diabetes diagnosis"; "sex classification"; and "diabetic retinopathy diagnosis".

Discussion

BRSET is the first multilabel ophthalmological dataset in Brazil and Latin America. It provides an opportunity for investigating model biases by evaluating performance across demographic groups. The model performance of three prediction tasks demonstrates the value of the dataset for external validation and for teaching medical computer vision to learners in Latin America using locally relevant data sources.

A scientometric analysis of fairness in health AI literature

Artificial intelligence (AI) and machine learning are central components of today's medical environment. The fairness of AI, i.e. the ability of AI to be free from bias, has repeatedly come into question. This study investigates the diversity of members of academia whose scholarship poses questions about the fairness of AI. The articles that combine the topics of fairness, artificial intelligence, and medicine were selected from Pubmed, Google Scholar, and Embase using keywords. Eligibility and data extraction from the articles were done manually and cross-checked by another author for accuracy. Articles were selected for further analysis, cleaned, and organized in Microsoft Excel; spatial diagrams were generated using Public Tableau. Additional graphs were generated using Matplotlib and Seaborn. Linear and logistic regressions were conducted using Python to measure the relationship between funding status, number of citations, and the gender demographics of the authorship team. We identified 375 eligible publications, including research and review articles concerning AI and fairness in healthcare. Analysis of the bibliographic data revealed that there is an overrepresentation of authors that are white, male, and are from high-income countries, especially in the roles of first and last author. Additionally, analysis showed that papers whose authors are based in higher-income countries were more likely to be cited more often and published in higher impact journals. These findings highlight the lack of diversity among the authors in the AI fairness community whose work gains the largest readership, potentially compromising the very impartiality that the AI fairness community is working towards.

Caregivers Attitude Detection From Clinical Notes

Caregivers' attitudes impact healthcare quality and disparities. Clinical notes contain highly specialized and ambiguous language that requires extensive domain knowledge to understand, and using negative language does not necessarily imply a negative attitude. This study discusses the challenge of detecting caregivers' attitudes from their clinical notes. To address these challenges, we annotate MIMIC clinical notes and train state-of-the-art language models from the Hugging Face platform. The study focuses on the Neonatal Intensive Care Unit and evaluates models in zero-shot, few-shot, and fully-trained scenarios. Among the chosen models, RoBERTa identifies caregivers' attitudes from clinical notes with an F1-score of 0.75. This approach not only enhances patient satisfaction, but opens up exciting possibilities for detecting and preventing care provider syndromes, such as fatigue, stress, and burnout. The paper concludes by discussing limitations and potential future work.

Peer review of GPT-4 technical report and systems card

The study provides a comprehensive review of OpenAI's Generative Pre-trained Transformer 4 (GPT-4) technical report, with an emphasis on applications in high-risk settings like healthcare. A diverse team, including experts in artificial intelligence (AI), natural language processing, public health, law, policy, social science, healthcare research, and bioethics, analyzed the report against established peer review guidelines. The GPT-4 report shows a significant commitment to transparent AI research, particularly in creating a systems card for risk assessment and mitigation. However, it reveals limitations such as restricted access to training data, inadequate confidence and uncertainty estimations, and concerns over privacy and intellectual property rights. Key strengths identified include the considerable time and economic investment in transparent AI research and the creation of a comprehensive systems card. On the other hand, the lack of clarity in training processes and data raises concerns about encoded biases and interests in GPT-4. The report also lacks confidence and uncertainty estimations, crucial in high-risk areas like healthcare, and fails to address potential privacy and intellectual property issues. Furthermore, this study emphasizes the need for diverse, global involvement in developing and evaluating large language models (LLMs) to ensure broad societal benefits and mitigate risks. The paper presents recommendations such as improving data transparency, developing accountability frameworks, establishing confidence standards for LLM outputs in high-risk settings, and enhancing industry research review processes. It concludes that while GPT-4's report is a step towards open discussions on LLMs, more extensive interdisciplinary reviews are essential for addressing bias, harm, and risk concerns, especially in high-risk domains. The review aims to expand the understanding of LLMs in general and highlights the need for new reflection forms on how LLMs are reviewed, the data required for effective evaluation, and addressing critical issues like bias and risk.

An introduction to digital determinants of health

In recent years, technology has been increasingly incorporated within healthcare for the provision of safe and efficient delivery of services. Although this can be attributed to the benefits that can be harnessed, digital technology has the potential to exacerbate and reinforce preexisting health disparities. Previous work has highlighted how sociodemographic, economic, and political factors affect individuals' interactions with digital health systems and are termed social determinants of health [SDOH]. But, there is a paucity of literature addressing how the intrinsic design, implementation, and use of technology interact with SDOH to influence health outcomes. Such interactions are termed digital determinants of health [DDOH]. This paper will, for the first time, propose a definition of DDOH and provide a conceptual model characterizing its influence on healthcare outcomes. Specifically, DDOH is implicit in the design of artificial intelligence systems, mobile phone applications, telemedicine, digital health literacy [DHL], and other forms of digital technology. A better appreciation of DDOH by the various stakeholders at the individual and societal levels can be channeled towards policies that are more digitally inclusive. In tandem with ongoing work to minimize the digital divide caused by existing SDOH, further work is necessary to recognize digital determinants as an important and distinct entity.

Challenges in the maintenance of an open hospital-based cancer registry system in a low-to-middle-income country (LMIC): 2017-2022 experience

Hospital-based cancer registries (HBCRs) record data on all patients diagnosed and/or treated for cancer at healthcare facilities and evaluate the burden of the disease and the quality of healthcare services at that hospital, helping improve patient care, and providing an assessment of healthcare quality. The CARE PH app was created as a tool to facilitate a system of hospital-based cancer registries in the Philippines, a lower middle-income country. From 2017 to 2022, a total of 60,021 cancer registrants from 44 CARE PH hospitals were entered into the database. Breast cancer was the most common primary site, accounting for 17,660 cases (29.4%). This was followed by colorectal cancer at 11.1%, cervical cancer at 6.2%, head and neck cancer at 5.9%, and prostate and other male genital cancer at 5.1%.Among the 30 data fields collected, 17 exhibited 0-20% missing data, eight displayed 21%-90% missing data, while five depicted 91%-100% missing data. Most of the data fields with missing data are in the treatment and follow-up modules, which are stored in separate forms in a patient's record. Digital transformation of hospitals from paper-based charts to electronic medical records, and the integration of the HBCR to the EMR and hospital information system, will likely be the best solution for these limitations. It is recommended that the creation and maintenance of HBCRs nationwide must be harmonized, and embedded in all relevant national programs and legislations. The development of an information technology process that is based on a cancer patient's journey, should be built on an open system embedded in a well designed enterprise architecture, functioning under the guidance of a strong leadership and governance team. All these must be present in order to create and maintain a robust HBCR that is useful for furthering cancer registry and research in the country.

Assessing the potential of GPT-4 to perpetuate racial and gender biases in health care: a model evaluation study

BACKGROUND

Large language models (LLMs) such as GPT-4 hold great promise as transformative tools in health care, ranging from automating administrative tasks to augmenting clinical decision making. However, these models also pose a danger of perpetuating biases and delivering incorrect medical diagnoses, which can have a direct, harmful impact on medical care. We aimed to assess whether GPT-4 encodes racial and gender biases that impact its use in health care.

METHODS

Using the Azure OpenAI application interface, this model evaluation study tested whether GPT-4 encodes racial and gender biases and examined the impact of such biases on four potential applications of LLMs in the clinical domain-namely, medical education, diagnostic reasoning, clinical plan generation, and subjective patient assessment. We conducted experiments with prompts designed to resemble typical use of GPT-4 within clinical and medical education applications. We used clinical vignettes from NEJM Healer and from published research on implicit bias in health care. GPT-4 estimates of the demographic distribution of medical conditions were compared with true US prevalence estimates. Differential diagnosis and treatment planning were evaluated across demographic groups using standard statistical tests for significance between groups.

FINDINGS

We found that GPT-4 did not appropriately model the demographic diversity of medical conditions, consistently producing clinical vignettes that stereotype demographic presentations. The differential diagnoses created by GPT-4 for standardised clinical vignettes were more likely to include diagnoses that stereotype certain races, ethnicities, and genders. Assessment and plans created by the model showed significant association between demographic attributes and recommendations for more expensive procedures as well as differences in patient perception.

INTERPRETATION

Our findings highlight the urgent need for comprehensive and transparent bias assessments of LLM tools such as GPT-4 for intended use cases before they are integrated into clinical care. We discuss the potential sources of these biases and potential mitigation strategies before clinical implementation.

FUNDING

Priscilla Chan and Mark Zuckerberg.

Evaluating Prognostic Bias of Critical Illness Severity Scores Based on Age, Sex, and Primary Language in the United States: A Retrospective Multicenter Study

OBJECTIVES

Although illness severity scoring systems are widely used to support clinical decision-making and assess ICU performance, their potential bias across different age, sex, and primary language groups has not been well-studied.

DESIGN SETTING AND PATIENTS

We aimed to identify potential bias of Sequential Organ Failure Assessment (SOFA) and Acute Physiology and Chronic Health Evaluation (APACHE) IVa scores via large ICU databases.

SETTING/PATIENTS

This multicenter, retrospective study was conducted using data from the Medical Information Mart for Intensive Care (MIMIC) and eICU Collaborative Research Database. SOFA and APACHE IVa scores were obtained from ICU admission. Hospital mortality was the primary outcome. Discrimination (area under receiver operating characteristic [AUROC] curve) and calibration (standardized mortality ratio [SMR]) were assessed for all subgroups.

INTERVENTIONS

Not applicable.

MEASUREMENTS AND MAIN RESULTS

A total of 196,310 patient encounters were studied. Discrimination for both scores was worse in older patients compared with younger patients and female patients rather than male patients. In MIMIC, discrimination of SOFA in non-English primary language speakers patients was worse than that of English speakers (AUROC 0.726 vs. 0.783, p < 0.0001). Evaluating calibration via SMR showed statistically significant underestimations of mortality when compared with overall cohort in the oldest patients for both SOFA and APACHE IVa, female patients (1.09) for SOFA, and non-English primary language patients (1.38) for SOFA in MIMIC.

CONCLUSIONS

Differences in discrimination and calibration of two scores across varying age, sex, and primary language groups suggest illness severity scores are prone to bias in mortality predictions. Caution must be taken when using them for quality benchmarking and decision-making among diverse real-world populations.

Interdisciplinary collaboration in critical care alarm research: A bibliometric analysis

BACKGROUND

Alarm fatigue in nurses is a major patient safety concern in the intensive care unit. This is caused by exposure to high rates of false and non-actionable alarms. Despite decades of research, the problem persists, leading to stress, burnout, and patient harm resulting from true missed events. While engineering approaches to reduce false alarms have spurred hope, they appear to lack collaboration between nurses and engineers to produce real-world solutions. The aim of this bibliometric analysis was to examine the relevant literature to quantify the level of authorial collaboration between nurses, physicians, and engineers.

METHODS

We conducted a bibliometric analysis of articles on alarm fatigue and false alarm reduction strategies in critical care published between 2010 and 2022. Data were extracted at the article and author level. The percentages of author disciplines per publication were calculated by study design, journal subject area, and other article-level factors.

RESULTS

A total of 155 articles with 583 unique authors were identified. While 31.73 % (n = 185) of the unique authors had a nursing background, publications using an engineering study design (n = 46), e.g., model development, had a very low involvement of nursing authors (mean proportion at 1.09 %). Observational studies (n = 58) and interventional studies (n = 33) had a higher mean involvement of 52.27 % and 47.75 %, respectively. Articles published in nursing journals (n = 32) had the highest mean proportion of nursing authors (80.32 %), while those published in engineering journals (n = 46) had the lowest (9.00 %), with 6 (13.04 %) articles having one or more nurses as co-authors.

CONCLUSION

Minimal involvement of nursing expertise in alarm research utilizing engineering methodologies may be one reason for the lack of successful, real-world solutions to ameliorate alarm fatigue. Fostering a collaborative, interdisciplinary research culture can promote a common publication culture across fields and may yield sustainable implementation of technological solutions in healthcare.

Investigating the Impact of Automation on the Health Care Workforce Through Autonomous Telemedicine in the Cataract Pathway: Protocol for a Multicenter Study

BACKGROUND

While digital health innovations are increasingly being adopted by health care organizations, implementation is often carried out without considering the impacts on frontline staff who will be using the technology and who will be affected by its introduction. The enthusiasm surrounding the use of artificial intelligence (AI)-enabled digital solutions in health care is tempered by uncertainty around how it will change the working lives and practices of health care professionals. Digital enablement can be viewed as facilitating enhanced effectiveness and efficiency by improving services and automating cognitive labor, yet the implementation of such AI technology comes with challenges related to changes in work practices brought by automation. This research explores staff experiences before and after care pathway automation with an autonomous clinical conversational assistant, Dora (Ufonia Ltd), that is able to automate routine clinical conversations.

OBJECTIVE

The primary objective is to examine the impact of AI-enabled automation on clinicians, allied health professionals, and administrators who provide or facilitate health care to patients in high-volume, low-complexity care pathways. In the process of transforming care pathways through automation of routine tasks, staff will increasingly "work at the top of their license." The impact of this fundamental change on the professional identity, well-being, and work practices of the individual is poorly understood at present.

METHODS

We will adopt a multiple case study approach, combining qualitative and quantitative data collection methods, over 2 distinct phases, namely phase A (preimplementation) and phase B (postimplementation).

RESULTS

The analysis is expected to reveal the interrelationship between Dora and those affected by its introduction. This will reveal how tasks and responsibilities have changed or shifted, current tensions and contradictions, ways of working, and challenges, benefits, and opportunities as perceived by those on the frontlines of the health care system. The findings will enable a better understanding of the resistance or susceptibility of different stakeholders within the health care workforce and encourage managerial awareness of differing needs, demands, and uncertainties.

CONCLUSIONS

The implementation of AI in the health care sector, as well as the body of research on this topic, remain in their infancy. The project's key contribution will be to understand the impact of AI-enabled automation on the health care workforce and their work practices.

INTERNATIONAL REGISTERED REPORT IDENTIFIER (IRRID)

PRR1-10.2196/49374.

VAP risk index: Early prediction and hospital phenotyping of ventilator-associated pneumonia using machine learning

BACKGROUND

Ventilator-associated pneumonia (VAP) is a leading cause of morbidity and mortality in intensive care units (ICUs). Early identification of patients at risk of VAP enables early intervention, which in turn improves patient outcomes. We developed a predictive model for individualized risk assessment utilizing machine learning to identify patients at risk of developing VAP.

METHODS

The Philips eRI dataset, a multi-institution electronic medical record (EMR), was used for model development. For adult (≥18y) patients, we propose a set of criteria using indications of the start of a new antibiotic treatment temporally contiguous to a microbiological test to mark suspected infection events, of which those with a positive culture are labeled as presumed VAP if 1) the event occurs at least 48 h after intubation, and 2) there are no indications of community-acquired pneumonia (CAP) or other hospital-acquired infections (HAI) in the patient charts. The resulting VAP and no-VAP (control) cases were then used to build an ensemble of decision trees to predict the risk of VAP in the next 24 h using data on patients' demographics, vitals, labs, and ventilator settings.

RESULTS

The resulting model predicts the development of VAP 24 h in advance with an AUC of 76 % and AUPRC of 75 %. Additionally, we group hospitals that are similar in healthcare processes into distinct clusters and characterize VAP prediction for the identified hospital clusters. We show inter-hospital (teaching status and healthcare processes) and cohort-specific (age groups, gender, early vs late VAP, ICU mortality status) differences in VAP prediction and associated symptomologies.

CONCLUSIONS

Our proposed VAP criteria use clinical actions to mark incidences of presumed VAP infection, which enables the development of models for early detection of these events. We curated a patient cohort using these criteria and used it to build a model for predicting impending VAP events prior to clinical suspicions. We present a clustering approach for tailoring the VAP prediction model for different hospital types based on their EMR data characteristics. The model provides an instantaneous risk score that allows early interventions and confirmatory diagnostic actions.

Standardized Comparison of Voice-Based Information and Documentation Systems to Established Systems in Intensive Care: Crossover Study

BACKGROUND

The medical teams in intensive care units (ICUs) spend increasing amounts of time at computer systems for data processing, input, and interpretation purposes. As each patient creates about 1000 data points per hour, the available information is abundant, making the interpretation difficult and time-consuming. This data flood leads to a decrease in time for evidence-based, patient-centered care. Information systems, such as patient data management systems (PDMSs), are increasingly used at ICUs. However, they often create new challenges arising from the increasing documentation burden.

OBJECTIVE

New concepts, such as artificial intelligence (AI)-based assistant systems, are hence introduced to the workflow to cope with these challenges. However, there is a lack of standardized, published metrics in order to compare the various data input and management systems in the ICU setting. The objective of this study is to compare established documentation and retrieval processes with newer methods, such as PDMSs and voice information and documentation systems (VIDSs).

METHODS

In this crossover study, we compare traditional, paper-based documentation systems with PDMSs and newer AI-based VIDSs in terms of performance (required time), accuracy, mental workload, and user experience in an intensive care setting. Performance is assessed on a set of 6 standardized, typical ICU tasks, ranging from documentation to medical interpretation.

RESULTS

A total of 60 ICU-experienced medical professionals participated in the study. The VIDS showed a statistically significant advantage compared to the other 2 systems. The tasks were completed significantly faster with the VIDS than with the PDMS (1-tailed t59=12.48; Cohen d=1.61; P<.001) or paper documentation (t59=20.41; Cohen d=2.63; P<.001). Significantly fewer errors were made with VIDS than with the PDMS (t59=3.45; Cohen d=0.45; P=.03) and paper-based documentation (t59=11.2; Cohen d=1.45; P<.001). The analysis of the mental workload of VIDS and PDMS showed no statistically significant difference (P=.06). However, the analysis of subjective user perception showed a statistically significant perceived benefit of the VIDS compared to the PDMS (P<.001) and paper documentation (P<.001).

CONCLUSIONS

The results of this study show that the VIDS reduced error rate, documentation time, and mental workload regarding the set of 6 standardized typical ICU tasks. In conclusion, this indicates that AI-based systems such as the VIDS tested in this study have the potential to reduce this workload and improve evidence-based and safe patient care.

Disparity dashboards: an evaluation of the literature and framework for health equity improvement

The growing recognition of differences in health outcomes across populations has led to a slow but increasing shift towards transparent reporting of patient outcomes. In addition, pay-for-equity initiatives, such as those proposed by the Centers for Medicare and Medicaid, will require the reporting of health outcomes across subgroups over time. Dashboards offer one means of visualising data in the health-care context that can highlight essential disparities in clinical outcomes, guide targeted quality-improvement efforts, and ultimately improve health equity. In this Viewpoint, we evaluate all studies that have reported the successful development of a disparity dashboard and share the data collected and unintended consequences reported. We propose a framework for systematic equality improvement through incentivisation of the collecting and reporting of health data and through implementation of reward systems to reduce health disparities.

Patient diversity and author representation in clinical studies supporting the Surviving Sepsis Campaign guidelines for management of sepsis and septic shock 2021: a systematic review of citations

BACKGROUND

The generalizability of the Surviving Sepsis Campaign (SSC) guidelines to various patient populations and hospital settings has been debated. A quantitative assessment of the diversity and representation in the clinical evidence supporting the guidelines would help evaluate the generalizability of the recommendations and identify strategic research goals and priorities. In this study, we evaluated the diversity of patients in the original studies, in terms of sex, race/ethnicity, and geographical location. We also assessed diversity in sex and geographical representation among study first and last authors.

METHODS

All clinical studies cited in support of the 2021 SSC adult guideline recommendations were identified. Original clinical studies were included, while editorials, reviews, non-clinical studies, and meta-analyses were excluded. For eligible studies, we recorded the proportion of male patients, percentage of each represented racial/ethnic subgroup (when available), and countries in which they were conducted. We also recorded the sex and location of the first and last authors. The World Bank classification was used to categorize countries.

RESULTS

The SSC guidelines included six sections, with 85 recommendations based on 351 clinical studies. The proportion of male patients ranged from 47 to 62%. Most studies did not report the racial/ ethnic distribution of the included patients; when they did so, most were White patients (68-77%). Most studies were conducted in high-income countries (77-99%), which included Europe/Central Asia (33-66%) and North America (36-55%). Moreover, most first/last authors were males (55-93%) and from high-income countries (77-99%).

CONCLUSIONS

To enhance the generalizability of the SCC guidelines, stakeholders should define strategies to enhance the diversity and representation in clinical studies. Though there was reasonable representation in sex among patients included in clinical studies, the evidence did not reflect diversity in the race/ethnicity and geographical locations. There was also lack of diversity among the first and last authors contributing to the evidence.

Exploring disease axes as an alternative to distinct clusters for characterizing sepsis heterogeneity

PURPOSE

Various studies have analyzed sepsis subtypes, yet the reproducibility of such results remains unclear. This study aimed to determine the reproducibility of sepsis subtypes across multiple cohorts.

METHODS

The study examined 63,547 sepsis patients from six distinct cohorts who had similar sepsis-related characteristics (vital signs, lactate, sequential organ failure assessment score, bilirubin, serum, urine output, and Glasgow coma scale). Identical cluster analysis techniques were used, employing 27 clustering schemes, and normalized mutual information (NMI), a metric ranging from 0 to 1 with higher values indicating better concordance, was employed to quantify the clustering solutions' reproducibility. Principal component analysis (PCA) was utilized to obtain the disease axis, and its uniformity across cohorts was evaluated through patterns of feature loading and correlation.

RESULTS

The reproducibility of sepsis clustering subtypes across the various studies was modest (median NMI ranging from 0.08 to 0.54). The top-down transfer learning method (model trained on cohorts with greater severity was transferred to cohorts with lower severity score) had a higher NMI value than the bottom-up approach (median [Q1, Q3]: 0.64 [0.49, 0.78] vs. 0.23 [0.2, 0.31], p < 0.001). The reproducibility was greater when the transfer solution was performed within United States (US) cohorts. The PCA analysis revealed that the correlation pattern between variables was consistent across all cohorts, and the first two disease axes were the "shock axis" and "systemic inflammatory response syndrome (SIRS) axis."

CONCLUSIONS

Cluster analysis of sepsis patients across various cohorts showed modest reproducibility. Sepsis heterogeneity is better characterized through continuous disease axes that coexist to varying degrees within the same individual instead of mutually exclusive subtypes.

From quality improvement to equality improvement projects: A scoping review and framework

Increasing awareness of health disparities has led to proposals for a pay-for-equity scheme. Implementing such proposals requires systematic methods of collecting and reporting health outcomes for targeted demographics over time. This lays the foundation for a shift from quality improvement projects (QIPs) to equality improvement projects (EQIPs) that could evaluate adherence to standards and progress toward health equity. We performed a scoping review on EQIPs to inform a new framework for quality improvement through a health equity lens. Forty studies implemented an intervention after identifying a disparity compared to 149 others which merely identified group differences. Most evaluated race-based differences and were conducted at the institutional level, with representation in both the inpatient and outpatient settings. EQIPs that improved equity leveraged multidisciplinary expertise, healthcare staff education, and developed tools to track health outcomes continuously. EQIPs can help bridge the inequality gap and form part of an incentivized systematic equality improvement framework.

Variation in monitoring: Glucose measurement in the ICU as a case study to preempt spurious correlations

Objective

Health inequities can be influenced by demographic factors such as race and ethnicity, proficiency in English, and biological sex. Disparities may manifest as differential likelihood of testing which correlates directly with the likelihood of an intervention to address an abnormal finding. Our retrospective observational study evaluated the presence of variation in glucose measurements in the Intensive Care Unit (ICU).

Methods

Using the MIMIC-IV database (2008-2019), a single-center, academic referral hospital in Boston (USA), we identified adult patients meeting sepsis-3 criteria. Exclusion criteria were diabetic ketoacidosis, ICU length of stay under 1 day, and unknown race or ethnicity. We performed a logistic regression analysis to assess differential likelihoods of glucose measurements on day 1. A negative binomial regression was fitted to assess the frequency of subsequent glucose readings. Analyses were adjusted for relevant clinical confounders, and performed across three disparity proxy axes: race and ethnicity, sex, and English proficiency.

Results

We studied 24,927 patients, of which 19.5% represented racial and ethnic minority groups, 42.4% were female, and 9.8% had limited English proficiency. No significant differences were found for glucose measurement on day 1 in the ICU. This pattern was consistent irrespective of the axis of analysis, i.e. race and ethnicity, sex, or English proficiency. Conversely, subsequent measurement frequency revealed potential disparities. Specifically, males (incidence rate ratio (IRR) 1.06, 95% confidence interval (CI) 1.01 - 1.21), patients who identify themselves as Hispanic (IRR 1.11, 95% CI 1.01 - 1.21), or Black (IRR 1.06, 95% CI 1.01 - 1.12), and patients being English proficient (IRR 1.08, 95% CI 1.01 - 1.15) had higher chances of subsequent glucose readings.

Conclusion

We found disparities in ICU glucose measurements among patients with sepsis, albeit the magnitude was small. Variation in disease monitoring is a source of data bias that may lead to spurious correlations when modeling health data.

Evaluating equitable care in the ICU: Creating a causal inference framework to assess the impact of life-sustaining interventions across racial and ethnic groups

Background

Variability in the provision of intensive care unit (ICU)-interventions may lead to disparities between socially defined racial-ethnic groups.

Research Question

We used causal inference to examine the use of invasive mechanical ventilation (IMV), renal replacement therapy (RRT), and vasopressor agents (VP) to identify disparities in outcomes across race-ethnicity in patients with sepsis.

Study Design and Methods

Single-center, academic referral hospital in Boston, Massachusetts, USA. Retrospective analysis of treatment effect with a targeted trial design categorized by treatment assignment within the first 24 hours in the MIMIC-IV dataset (2008- 2019) using targeted maximum likelihood estimation. Of 76,943 ICU stays in MIMIC-IV, 32,971 adult stays fulfilling sepsis-3 criteria were included. The primary outcome was in-hospital mortality. Secondary outcomes were hospital-free days, and occurrence of nosocomial infection stratified by predicted mortality probability ranges and self-reported race-ethnicity. Average treatment effects by treatment type and race-ethnicity, Racial-ethnic group (REG) or White group (WG), were estimated.

Results

Of 19,419 admissions that met inclusion criteria, median age was 68 years, 57.4% were women, 82% were White, and mortality was 18.2%. There was no difference in mortality benefit associated with the administration of IMV, RRT, or VP between the REG and the WG. There was also no difference in hospital-free days or nosocomial infections. These findings are unchanged with different eligibility periods.

Interpretation

There were no differences in the treatment outcomes from three life-sustaining interventions in the ICU according to race-ethnicity. While there was no discernable harm from the treatments across mortality risk, there was also no measurable benefit. These findings highlight the need for research to understand better the risk-benefit of life-sustaining interventions in the ICU.

BOLD: Blood-gas and Oximetry Linked Dataset - Open Source Research

Pulse oximeters measure peripheral arterial oxygen saturation (SpO 2 ) noninvasively, while the gold standard (SaO 2 ) involves arterial blood gas measurement. There are known racial and ethnic disparities in their performance. BOLD is a new comprehensive dataset that aims to underscore the importance of addressing biases in pulse oximetry accuracy, which disproportionately affect darker-skinned patients. The dataset was created by harmonizing three Electronic Health Record databases (MIMIC-III, MIMIC-IV, eICU-CRD) comprising Intensive Care Unit stays of US patients. Paired SpO 2 and SaO 2 measurements were time-aligned and combined with various other sociodemographic and parameters to provide a detailed representation of each patient. BOLD includes 49,099 paired measurements, within a 5-minute window and with oxygen saturation levels between 70-100%. Minority racial and ethnic groups account for ∼25% of the data - a proportion seldom achieved in previous studies. The codebase is publicly available. Given the prevalent use of pulse oximeters in the hospital and at home, we hope that BOLD will be leveraged to develop debiasing algorithms that can result in more equitable healthcare solutions.

Illness severity assessment of older adults in critical illness using machine learning (ELDER-ICU): an international multicentre study with subgroup bias evaluation

BACKGROUND

Comorbidity, frailty, and decreased cognitive function lead to a higher risk of death in elderly patients (more than 65 years of age) during acute medical events. Early and accurate illness severity assessment can support appropriate decision making for clinicians caring for these patients. We aimed to develop ELDER-ICU, a machine learning model to assess the illness severity of older adults admitted to the intensive care unit (ICU) with cohort-specific calibration and evaluation for potential model bias.

METHODS

In this retrospective, international multicentre study, the ELDER-ICU model was developed using data from 14 US hospitals, and validated in 171 hospitals from the USA and Netherlands. Data were extracted from the Medical Information Mart for Intensive Care database, electronic ICU Collaborative Research Database, and Amsterdam University Medical Centers Database. We used six categories of data as predictors, including demographics and comorbidities, physical frailty, laboratory tests, vital signs, treatments, and urine output. Patient data from the first day of ICU stay were used to predict in-hospital mortality. We used the eXtreme Gradient Boosting algorithm (XGBoost) to develop models and the SHapley Additive exPlanations method to explain model prediction. The trained model was calibrated before internal, external, and temporal validation. The final XGBoost model was compared against three other machine learning algorithms and five clinical scores. We performed subgroup analysis based on age, sex, and race. We assessed the discrimination and calibration of models using the area under receiver operating characteristic (AUROC) and standardised mortality ratio (SMR) with 95% CIs.

FINDINGS

Using the development dataset (n=50 366) and predictive model building process, the XGBoost algorithm performed the best in all types of validations compared with other machine learning algorithms and clinical scores (internal validation with 5037 patients from 14 US hospitals, AUROC=0·866 [95% CI 0·851-0·880]; external validation in the US population with 20 541 patients from 169 hospitals, AUROC=0·838 [0·829-0·847]; external validation in European population with 2411 patients from one hospital, AUROC=0·833 [0·812-0·853]; temporal validation with 4311 patients from one hospital, AUROC=0·884 [0·869-0·897]). In the external validation set (US population), the median AUROCs of bias evaluations covering eight subgroups were above 0·81, and the overall SMR was 0·99 (0·96-1·03). The top ten risk predictors were the minimum Glasgow Coma Scale score, total urine output, average respiratory rate, mechanical ventilation use, best state of activity, Charlson Comorbidity Index score, geriatric nutritional risk index, code status, age, and maximum blood urea nitrogen. A simplified model containing only the top 20 features (ELDER-ICU-20) had similar predictive performance to the full model.

INTERPRETATION

The ELDER-ICU model reliably predicts the risk of in-hospital mortality using routinely collected clinical features. The predictions could inform clinicians about patients who are at elevated risk of deterioration. Prospective validation of this model in clinical practice and a process for continuous performance monitoring and model recalibration are needed.

FUNDING

National Institutes of Health, National Natural Science Foundation of China, National Special Health Science Program, Health Science and Technology Plan of Zhejiang Province, Fundamental Research Funds for the Central Universities, Drug Clinical Evaluate Research of Chinese Pharmaceutical Association, and National Key R&D Program of China.

Critical Bias in Critical Care Devices

Critical care data contain information about the most physiologically fragile patients in the hospital, who require a significant level of monitoring. However, medical devices used for patient monitoring suffer from measurement biases that have been largely underreported. This article explores sources of bias in commonly used clinical devices, including pulse oximeters, thermometers, and sphygmomanometers. Further, it provides a framework for mitigating these biases and key principles to achieve more equitable health care delivery.

Emulating Target Trials Comparing Early and Delayed Intubation Strategies

BACKGROUND

Whether intubation should be initiated early in the clinical course of critically ill patients remains a matter of debate. Results from prior observational studies are difficult to interpret because of avoidable flaws including immortal time bias, inappropriate eligibility criteria, and unrealistic treatment strategies.

RESEARCH QUESTION

Do treatment strategies that intubate patients early in the critical care admission improve 30-day survival compared with strategies that delay intubation?

STUDY DESIGN AND METHODS

We estimated the effect of strategies that require early intubation of critically ill patients compared with those that delay intubation. With data extracted from the Medical Information Mart for Intensive Care-IV database, we emulated three target trials, varying the flexibility of the treatment strategies and the baseline eligibility criteria.

RESULTS

Under unrealistically strict treatment strategies with broad eligibility criteria, the 30-day mortality risk was 7.1 percentage points higher for intubating early compared with delaying intubation (95% CI, 6.2-7.9). Risk differences were 0.4 (95% CI, -0.1 to 0.9) and -0.9 (95% CI, -2.5 to 0.7) percentage points in subsequent target trial emulations that included more realistic treatment strategies and eligibility criteria.

INTERPRETATION

When realistic treatment strategies and eligibility criteria are used, strategies that delay intubation result in similar 30-day mortality risks compared with those that intubate early. Delaying intubation ultimately avoids intubation in most patients.

A scoping review of the landscape of health-related open datasets in Latin America

Artificial intelligence (AI) algorithms have the potential to revolutionize healthcare, but their successful translation into clinical practice has been limited. One crucial factor is the data used to train these algorithms, which must be representative of the population. However, most healthcare databases are derived from high-income countries, leading to non-representative models and potentially exacerbating health inequities. This review focuses on the landscape of health-related open datasets in Latin America, aiming to identify existing datasets, examine data-sharing frameworks, techniques, platforms, and formats, and identify best practices in Latin America. The review found 61 datasets from 23 countries, with the DATASUS dataset from Brazil contributing to the majority of articles. The analysis revealed a dearth of datasets created by the authors themselves, indicating a reliance on existing open datasets. The findings underscore the importance of promoting open data in Latin America. We provide recommendations for enhancing data sharing in the region.

AI pitfalls and what not to do: mitigating bias in AI

Various forms of artificial intelligence (AI) applications are being deployed and used in many healthcare systems. As the use of these applications increases, we are learning the failures of these models and how they can perpetuate bias. With these new lessons, we need to prioritize bias evaluation and mitigation for radiology applications; all the while not ignoring the impact of changes in the larger enterprise AI deployment which may have downstream impact on performance of AI models. In this paper, we provide an updated review of known pitfalls causing AI bias and discuss strategies for mitigating these biases within the context of AI deployment in the larger healthcare enterprise. We describe these pitfalls by framing them in the larger AI lifecycle from problem definition, data set selection and curation, model training and deployment emphasizing that bias exists across a spectrum and is a sequela of a combination of both human and machine factors.

Strategies and solutions to address Digital Determinants of Health (DDOH) across underinvested communities

Healthcare has long struggled to improve services through technology without further widening health disparities. With the significant expansion of digital health, a group of healthcare professionals and scholars from across the globe are proposing the official usage of the term "Digital Determinants of Health" (DDOH) to explicitly call out the relationship between technology, healthcare, and equity. This is the final paper in a series published in PLOS Digital Health that seeks to understand and summarize current knowledge of the strategies and solutions that help to mitigate the negative effects of DDOH for underinvested communities. Through a search of English-language Medline, Scopus, and Google Scholar articles published since 2010, 345 articles were identified that discussed the application of digital health technology among underinvested communities. A group of 8 reviewers assessed 132 articles selected at random for the mention of solutions that minimize differences in DDOH. Solutions were then organized by categories of policy; design and development; implementation and adoption; and evaluation and ongoing monitoring. The data were then assessed by category and the findings summarized. The reviewers also looked for common themes across the solutions and evidence of effectiveness. From this limited scoping review, the authors found numerous solutions mentioned across the papers for addressing DDOH and many common themes emerged regardless of the specific community or digital health technology under review. There was notably less information on solutions regarding ongoing evaluation and monitoring which corresponded with a lack of research evidence regarding effectiveness. The findings directionally suggest that universal strategies and solutions can be developed to address DDOH independent of the specific community under focus. With the need for the further development of DDOH measures, we also provide a framework for DDOH assessment.

An interpretable deep learning model for time-series electronic health records: Case study of delirium prediction in critical care

Deep Learning (DL) models have received increasing attention in the clinical setting, particularly in intensive care units (ICU). In this context, the interpretability of the outcomes estimated by the DL models is an essential step towards increasing adoption of DL models in clinical practice. To address this challenge, we propose an ante-hoc, interpretable neural network model. Our proposed model, named double self-attention architecture (DSA), uses two attention-based mechanisms, including self-attention and effective attention. It can capture the importance of input variables in general, as well as changes in importance along the time dimension for the outcome of interest. We evaluated our model using two real-world clinical datasets covering 22840 patients in predicting onset of delirium 12 h and 48 h in advance. Additionally, we compare the descriptive performance of our model with three post-hoc interpretable algorithms as well as with the opinion of clinicians based on the published literature and clinical experience. We find that our model covers the majority of the top-10 variables ranked by the other three post-hoc interpretable algorithms as well as the clinical opinion, with the advantage of taking into account both, the dependencies among variables as well as dependencies between varying time-steps. Finally, our results show that our model can improve descriptive performance without sacrificing predictive performance.

The promise of digital healthcare technologies

Digital health technologies have been in use for many years in a wide spectrum of healthcare scenarios. This narrative review outlines the current use and the future strategies and significance of digital health technologies in modern healthcare applications. It covers the current state of the scientific field (delineating major strengths, limitations, and applications) and envisions the future impact of relevant emerging key technologies. Furthermore, we attempt to provide recommendations for innovative approaches that would accelerate and benefit the research, translation and utilization of digital health technologies.

Artificial intelligence for telemedicine diabetic retinopathy screening: a review

PURPOSE

This study aims to compare artificial intelligence (AI) systems applied in diabetic retinopathy (DR) teleophthalmology screening, currently deployed systems, fairness initiatives and the challenges for implementation.

METHODS

The review included articles retrieved from PubMed/Medline/EMBASE literature search strategy regarding telemedicine, DR and AI. The screening criteria included human articles in English, Portuguese or Spanish and related to telemedicine and AI for DR screening. The author's affiliations and the study's population income group were classified according to the World Bank Country and Lending Groups.

RESULTS

The literature search yielded a total of 132 articles, and nine were included after full-text assessment. The selected articles were published between 2004 and 2020 and were grouped as telemedicine systems, algorithms, economic analysis and image quality assessment. Four telemedicine systems that perform a quality assessment, image preprocessing and pathological screening were reviewed. A data and post-deployment bias assessment are not performed in any of the algorithms, and none of the studies evaluate the social impact implementations. There is a lack of representativeness in the reviewed articles, with most authors and target populations from high-income countries and no low-income country representation.

CONCLUSIONS

Telemedicine and AI hold great promise for augmenting decision-making in medical care, expanding patient access and enhancing cost-effectiveness. Economic studies and social science analysis are crucial to support the implementation of AI in teleophthalmology screening programs. Promoting fairness and generalizability in automated systems combined with telemedicine screening programs is not straightforward. Improving data representativeness, reducing biases and promoting equity in deployment and post-deployment studies are all critical steps in model development.

A translational perspective towards clinical AI fairness

Artificial intelligence (AI) has demonstrated the ability to extract insights from data, but the fairness of such data-driven insights remains a concern in high-stakes fields. Despite extensive developments, issues of AI fairness in clinical contexts have not been adequately addressed. A fair model is normally expected to perform equally across subgroups defined by sensitive variables (e.g., age, gender/sex, race/ethnicity, socio-economic status, etc.). Various fairness measurements have been developed to detect differences between subgroups as evidence of bias, and bias mitigation methods are designed to reduce the differences detected. This perspective of fairness, however, is misaligned with some key considerations in clinical contexts. The set of sensitive variables used in healthcare applications must be carefully examined for relevance and justified by clear clinical motivations. In addition, clinical AI fairness should closely investigate the ethical implications of fairness measurements (e.g., potential conflicts between group- and individual-level fairness) to select suitable and objective metrics. Generally defining AI fairness as "equality" is not necessarily reasonable in clinical settings, as differences may have clinical justifications and do not indicate biases. Instead, "equity" would be an appropriate objective of clinical AI fairness. Moreover, clinical feedback is essential to developing fair and well-performing AI models, and efforts should be made to actively involve clinicians in the process. The adaptation of AI fairness towards healthcare is not self-evident due to misalignments between technical developments and clinical considerations. Multidisciplinary collaboration between AI researchers, clinicians, and ethicists is necessary to bridge the gap and translate AI fairness into real-life benefits.