Bibliometric analysis of the 3-year trends (2018-2021) in literature on artificial intelligence in ophthalmology and vision sciences.

OBJECTIVES: The objective of this analysis is to present a current view of the field of ophthalmology and vision research and artificial intelligence (AI) from topical and geographical perspectives. This will clarify the direction of the field in the future and aid clinicians in adapting to new technological developments. METHODS: A comprehensive search of four different databases was conducted. Statistical and bibliometric analysis were done to characterise the literature. Softwares used included the R Studio bibliometrix package, and VOSviewer. RESULTS: A total of 3939 articles were included in the final bibliometric analysis. Diabetic retinopathy (391, 6% of the top 100 keywords) was the most frequently occurring indexed keyword by a large margin. The highest impact literature was produced by the least populated countries and in those countries who collaborate internationally. This was confirmed via a hypothesis test where no correlation was found between gross number of published articles and average number of citations (p value=0.866, r=0.038), while graphing ratio of international collaboration against average citations produced a positive correlation (r=0.283). Majority of publications were found to be concentrated in journals specialising in vision and computer science, with this category of journals having the highest number of publications per journal (18.00 publications/journal), though they represented a small proportion of the total journals (<1%). CONCLUSION: This study provides a unique characterisation of the literature at the intersection of AI and ophthalmology and presents correlations between article impact and geography, in addition to summarising popular research topics.

Advancing equity in breast cancer care: natural language processing for analysing treatment outcomes in under-represented populations.

OBJECTIVE: The study aimed to develop natural language processing (NLP) algorithms to automate extracting patient-centred breast cancer treatment outcomes from clinical notes in electronic health records (EHRs), particularly for women from under-represented populations. METHODS: The study used clinical notes from 2010 to 2021 from a tertiary hospital in the USA. The notes were processed through various NLP techniques, including vectorisation methods (term frequency-inverse document frequency (TF-IDF), Word2Vec, Doc2Vec) and classification models (support vector classification, K-nearest neighbours (KNN), random forest (RF)). Feature selection and optimisation through random search and fivefold cross-validation were also conducted. RESULTS: The study annotated 100 out of 1000 clinical notes, using 970 notes to build the text corpus. TF-IDF and Doc2Vec combined with RF showed the highest performance, while Word2Vec was less effective. RF classifier demonstrated the best performance, although with lower recall rates, suggesting more false negatives. KNN showed lower recall due to its sensitivity to data noise. DISCUSSION: The study highlights the significance of using NLP in analysing clinical notes to understand breast cancer treatment outcomes in under-represented populations. The TF-IDF and Doc2Vec models were more effective in capturing relevant information than Word2Vec. The study observed lower recall rates in RF models, attributed to the dataset's imbalanced nature and the complexity of clinical notes. CONCLUSION: The study developed high-performing NLP pipeline to capture treatment outcomes for breast cancer in under-represented populations, demonstrating the importance of document-level vectorisation and ensemble methods in clinical notes analysis. The findings provide insights for more equitable healthcare strategies and show the potential for broader NLP applications in clinical settings.

Explainable machine learning for breast cancer diagnosis from mammography and ultrasound images: a systematic review.

BACKGROUND: Breast cancer is the most common disease in women. Recently, explainable artificial intelligence (XAI) approaches have been dedicated to investigate breast cancer. An overwhelming study has been done on XAI for breast cancer. Therefore, this study aims to review an XAI for breast cancer diagnosis from mammography and ultrasound (US) images. We investigated how XAI methods for breast cancer diagnosis have been evaluated, the existing ethical challenges, research gaps, the XAI used and the relation between the accuracy and explainability of algorithms. METHODS: In this work, Preferred Reporting Items for Systematic Reviews and Meta-Analyses checklist and diagram were used. Peer-reviewed articles and conference proceedings from PubMed, IEEE Explore, ScienceDirect, Scopus and Google Scholar databases were searched. There is no stated date limit to filter the papers. The papers were searched on 19 September 2023, using various combinations of the search terms 'breast cancer', 'explainable', 'interpretable', 'machine learning', 'artificial intelligence' and 'XAI'. Rayyan online platform detected duplicates, inclusion and exclusion of papers. RESULTS: This study identified 14 primary studies employing XAI for breast cancer diagnosis from mammography and US images. Out of the selected 14 studies, only 1 research evaluated humans' confidence in using the XAI system-additionally, 92.86% of identified papers identified dataset and dataset-related issues as research gaps and future direction. The result showed that further research and evaluation are needed to determine the most effective XAI method for breast cancer. CONCLUSION: XAI is not conceded to increase users' and doctors' trust in the system. For the real-world application, effective and systematic evaluation of its trustworthiness in this scenario is lacking. PROSPERO REGISTRATION NUMBER: CRD42023458665.

Towards inclusive biodesign and innovation: lowering barriers to entry in medical device development through large language model tools.

In the following narrative review, we discuss the potential role of large language models (LLMs) in medical device innovation, specifically examples using generative pretrained transformer-4. Throughout the biodesign process, LLMs can offer prompt-driven insights, aiding problem identification, knowledge assimilation and decision-making. Intellectual property analysis, regulatory assessment and market analysis emerge as key LLM applications. Through case examples, we underscore LLMs' transformative ability to democratise information access and expertise, facilitating inclusive innovation in medical devices as well as its effectiveness with providing real-time, individualised feedback for innovators of all experience levels. By mitigating entry barriers, LLMs accelerate transformative advancements, fostering collaboration among established and emerging stakeholders.

IT risk management for medical devices in hospital IT networks: a catalogue of measures and indicators.

OBJECTIVES: Connecting medical devices to hospital IT networks can create threats that must be covered by IT risk management. In practice, implementing such risk management is not trivial because the IEC 80001-1, as the existing state-of-the-art, do not describe sufficiently concrete implementation measures or evaluation indicators. The aim of the present work was to develop and evaluate a catalogue of measures and indicators to help hospitals implement and evaluate risk management in accordance with IEC 80001-1. METHODS: We conducted a Delphi study with 22 experts. In the first round, we performed interviews to identify implementation measures and evaluation indicators using qualitative content analysis. In the second round, a quantitative experts' survey confirmed the results of the first survey round and identified relationships between the measures and indicators. Based on these results, we then developed a catalogue containing the identified measures and indicators. Finally, we performed a case study to verify the practicability of this catalogue. RESULTS: We developed and verified a catalogue of 49 measures and 18 indicators to help hospitals implement and evaluate risk management following IEC 80001-1. The case study confirmed the practicability of the catalogue. DISCUSSION: Compared with IEC 80001-1, our catalogue goes into further detail to offer hospitals a stepwise implementation and evaluation approach. However, the catalogue must be tested in further case studies and evaluated in terms of generalisation. CONCLUSIONS: The catalogue will enable hospitals to overcome recent difficulties in implementing and evaluating IT risk management for medical devices according to IEC 80001-1.

Cluster analysis of dietary patterns associated with colorectal cancer derived from a Moroccan case-control study.

INTRODUCTION: Colorectal cancer (CRC) is a global public health problem. There is strong indication that nutrition could be an important component of primary prevention. Dietary patterns are a powerful technique for understanding the relationship between diet and cancer varying across populations. OBJECTIVE: We used an unsupervised machine learning approach to cluster Moroccan dietary patterns associated with CRC. METHODS: The study was conducted based on the reported nutrition of CRC matched cases and controls including 1483 pairs. Baseline dietary intake was measured using a validated food-frequency questionnaire adapted to the Moroccan context. Food items were consolidated into 30 food groups reduced on 6 dimensions by principal component analysis (PCA). RESULTS: K-means method, applied in the PCA-subspace, identified two patterns: 'prudent pattern' (moderate consumption of almost all foods with a slight increase in fruits and vegetables) and a 'dangerous pattern' (vegetable oil, cake, chocolate, cheese, red meat, sugar and butter) with small variation between components and clusters. The student test showed a significant relationship between clusters and all food consumption except poultry. The simple logistic regression test showed that people who belong to the 'dangerous pattern' have a higher risk to develop CRC with an OR 1.59, 95% CI (1.37 to 1.38). CONCLUSION: The proposed algorithm applied to the CCR Nutrition database identified two dietary profiles associated with CRC: the 'dangerous pattern' and the 'prudent pattern'. The results of this study could contribute to recommendations for CRC preventive diet in the Moroccan population.

Study of decision tree algorithms: effects of air pollution on under five mortality in Ulaanbaatar.

OBJECTIVES: 4.2 million people die every year from many diseases due to air pollution. The WHO confirms that 92% of the world's population lives in areas where the air quality limit is exceeded. In 251 days of 2011, the concentration of fine particulate matter in Ulaanbaatar exceeded the permissible level by 62%-76%. According to the results of the research, the content of fine particles decreased by 37%-46% in 2019. Because it is harmful to the health of children, we aimed to show the effect of air pollution on the mortality through data mining. METHODS: In many countries, research is being conducted to generate effective knowledge from big data using data mining methods. So, we are working to introduce this method to the health sector of Mongolia. In this study, we used the decision tree algorithms. RESULTS: We collected data on air pollution and under five mortality for 2019-2022 in Ulaanbaatar and created the database, built the models using the algorithms, and compared the results with the Mongolian standard. If the average of PM10 in winter is higher than the concentration specified in the standard, the mortality rate is likely to be high. Mortality is likely to be high if the nitrogen dioxide tolerance is high in the spring. CONCLUSION: The accuracy of the models calculated by the C5.0 algorithm is higher than the determined by the CART algorithm, the sensitivity and specificity values are higher than 0.50, so the mortality rates are uniformly predicted and low mortality prevails.

Health professionals' routine practice documentation and its associated factors in a resource-limited setting: a cross-sectional study.

OBJECTIVES: Documenting routine practice is significant for better diagnosis, treatment, continuity of care and medicolegal issues. However, health professionals' routine practice documentation is poorly practised. Therefore, this study aimed to assess health professionals' routine practice documentation and associated factors in a resource-limited setting. METHODS: An institution-based cross-sectional study design was used from 24 March up to 19 April 2022. Stratified random sampling and a pretested self-administered questionnaire were used among 423 samples. Epi Info V.7.1 and STATA V.15 software were used for data entry and analysis, respectively. Descriptive statistics and a logistic regression model were employed to describe the study subjects and to measure the strength of association between dependent and independent variables, respectively. A variable with a p value of <0.2 in bivariate logistic regression was considered for multivariable logistic regression. In multivariable logistic regression, ORs with 95% CIs and a p value of <0.05 were considered to determine the strength of association between dependent and independent variables. RESULTS: Health professionals' documentation practice was 51.1% (95% CI: 48.64 to 53.1). Lack of motivation (adjusted OR (AOR): 0.41, 95% CI: 0.22 to 0.76), good knowledge (AOR: 1.35, 95% CI: 0.72 to 2.97), taking training (AOR: 4.18, 95% CI: 2.99 to 8.28), using electronic systems (AOR: 2.19, 95% CI: 1.36 to 3.28), availability of standard documentation tools (AOR: 2.45, 95% CI: 1.35 to 4.43) were statistically associated factors. CONCLUSIONS: Health professionals' documentation practice is good. Lack of motivation, good knowledge, taking training, using electronic systems and the availability of documentation tools were significant factors. Stakeholders should provide additional training, and encourage professionals to use an electronic system for documentation practices.

Digital health and care: emerging from pandemic times.

In 2020, we published an editorial about the massive disruption of health and care services caused by the COVID-19 pandemic and the rapid changes in digital service delivery, artificial intelligence and data sharing that were taking place at the time. Now, 3 years later, we describe how these developments have progressed since, reflect on lessons learnt and consider key challenges and opportunities ahead by reviewing significant developments reported in the literature. As before, the three key areas we consider are digital transformation of services, realising the potential of artificial intelligence and wise data sharing to facilitate learning health systems. We conclude that the field of digital health has rapidly matured during the pandemic, but there are still major sociotechnical, evaluation and trust challenges in the development and deployment of new digital services.

From measures to action: can integrating quality measures provide system-wide insights for quality improvement decision making?

BACKGROUND: Quality improvement decision makers are left to develop an understanding of quality within their healthcare system from a deluge of narrowly focused measures that reflect existing fragmentation in care and lack a clear method for triggering improvement. A one-to-one metric-to-improvement strategy is intractable and leads to unintended consequences. Although composite measures have been used and their limitations noted in the literature, what remains unknown is 'Can integrating multiple quality measures provide a systemic understanding of care quality across a healthcare system?' METHODS: We devised a four-part data-driven analytic strategy to determine if consistent insights exist about the differential utilisation of end-of-life care using up to eight publicly available end-of-life cancer care quality measures across National Cancer Institute and National Comprehensive Cancer Network-designated cancer hospitals/centres. We performed 92 experiments that included 28 correlation analyses, 4 principal component analyses, 6 parallel coordinate analyses with agglomerative hierarchical clustering across hospitals and 54 parallel coordinate analyses with agglomerative hierarchical clustering within each hospital. RESULTS: Across 54 centres, integrating quality measures provided no consistent insights across different integration analyses. In other words, we could not integrate quality measures to describe how the underlying quality constructs of interest-intensive care unit (ICU) visits, emergency department (ED) visits, palliative care use, lack of hospice, recent hospice, use of life-sustaining therapy, chemotherapy and advance care planning-are used relative to each other across patients. Quality measure calculations lack interconnection information to construct a story that provides insights about where, when or what care is provided to which patients. And yet, we posit and discuss why administrative claims data-used to calculate quality measures-do contain such interconnection information. CONCLUSION: While integrating quality measures does not provide systemic information, new systemic mathematical constructs designed to convey interconnection information can be developed from the same administrative claims data to support quality improvement decision making.

Routine health information use among healthcare providers in Ethiopia: a systematic review and meta-analysis.

INTRODUCTION: Healthcare policy formulation, programme planning, monitoring and evaluation, and healthcare service delivery as a whole are dependent on routinely generated health information in a healthcare setting. Several individual research articles on the utilisation of routine health information exist in Ethiopia; however, each of them revealed inconsistent findings. OBJECTIVE: The main aim of this review was to combine the magnitude of routine health information use and its determinants among healthcare providers in Ethiopia. METHODS: Databases and repositories such as PubMed, Global Health, Scopus, Embase, African journal online, Advanced Google Search and Google Scholar were searched from 20 to 26 August 2022. RESULT: A total of 890 articles were searched but only 23 articles were included. A total of 8662 (96.3%) participants were included in the studies. The pooled prevalence of routine health information use was found to be 53.7% with 95% CI (47.45% to 59.95%). Training (adjusted OR (AOR)=1.56, 95% CI (1.12 to 2.18)), competency related to data management (AOR=1.94, 95% CI (1.35 to 2.8)), availability of standard guideline (AOR=1.66, 95% CI (1.38 to 1.99)), supportive supervision (AOR=2.07, 95% CI (1.55 to 2.76)) and feedback (AOR=2.20, 95% CI (1.30 to 3.71)) were significantly associated with routine health information use among healthcare providers at p value≤0.05 with 95% CI. CONCLUSION: The use of routinely generated health information for evidence-based decision-making remains one of the most difficult problems in the health information system. The study's reviewers suggested that the appropriate health authorities in Ethiopia invest in enhancing the skills in using routinely generated health information. PROSPERO REGISTRATION NUMBER: CRD42022352647.

Surgical pit crew: initiative to optimise measurement and accountability for operating room turnover time.

BACKGROUND AND OBJECTIVES: Turnover time (TOT), defined as the time between surgical cases in the same operating room (OR), is often perceived to be lengthy without clear cause. With the aim of optimising and standardising OR turnover processes and decreasing TOT, we developed an innovative and staff-interactive TOT measurement method. METHODS: We divided TOT into task-based segments and created buttons on the electronic health record (EHR) default prelogin screen for appropriate staff workflows to collect more granular data. We created submeasures, including 'clean-up start', 'clean-up complete', 'set-up start' and 'room ready for patient', to calculate environmental services (EVS) response time, EVS cleaning time, room set-up response time, room set-up time and time to room accordingly. RESULTS: Since developing and implementing these workflows, measures have demonstrated excellent staff adoption. Median times of EVS response and cleaning have decreased significantly at our main hospital ORs and ambulatory surgery centre. CONCLUSION: OR delays are costly to hospital systems. TOT, in particular, has been recognised as a potential dissatisfier and cause of delay in the perioperative environment. Viewing TOT as one finite entity and not a series of necessary tasks by a variety of team members limits the possibility of critical assessment and improvement. By dividing the measurement of TOT into respective segments necessary to transition the room at the completion of one case to the onset of another, valuable insight was gained into the causes associated with turnover delays, which increased awareness and improved accountability of staff members to complete assigned tasks efficiently.

Implementer report: ICD-10 code F44.5 review for functional seizure disorder.

OBJECTIVE: The study aimed to measure the validity of International Classification of Diseases, 10th Edition (ICD-10) code F44.5 for functional seizure disorder (FSD) in the Veterans Affairs Connecticut Healthcare System electronic health record (VA EHR). METHODS: The study used an informatics search tool, a natural language processing algorithm and a chart review to validate FSD coding. RESULTS: The positive predictive value (PPV) for code F44.5 was calculated to be 44%. DISCUSSION: ICD-10 introduced a specific code for FSD to improve coding validity. However, results revealed a meager (44%) PPV for code F44.5. Evaluation of the low diagnostic precision of FSD identified inconsistencies in the ICD-10 and VA EHR systems. CONCLUSION: Information system improvements may increase the precision of diagnostic coding by clinicians. Specifically, the EHR problem list should include commonly used diagnostic codes and an appropriately curated ICD-10 term list for 'seizure disorder,' and a single ICD code for FSD should be classified under neurology and psychiatry.

Use of a community advisory board to build equitable algorithms for participation in clinical trials: a protocol paper for HoPeNET.

INTRODUCTION: Participation from racial and ethnic minorities in clinical trials has been burdened by issues surrounding mistrust and access to healthcare. There is emerging use of machine learning (ML) in clinical trial recruitment and evaluation. However, for individuals from groups who are recipients of societal biases, utilisation of ML can lead to the creation and use of biased algorithms. To minimise bias, the design of equitable ML tools that advance health equity could be guided by community engagement processes. The Howard University Partnership with the National Institutes of Health for Equitable Clinical Trial Participation for Racial/Ethnic Communities Underrepresented in Research (HoPeNET) seeks to create an ML-based infrastructure from community advisory board (CAB) experiences to enhance participation of African-Americans/Blacks in clinical trials. METHODS AND ANALYSIS: This triphased cross-sectional study (24 months, n=56) will create a CAB of community members and research investigators. The three phases of the study include: (1) identification of perceived barriers/facilitators to clinical trial engagement through qualitative/quantitative methods and systems-based model building participation; (2) operation of CAB meetings and (3) development of a predictive ML tool and outcome evaluation. Identified predictors from the participant-derived systems-based map will be used for the ML tool development. ETHICS AND DISSEMINATION: We anticipate minimum risk for participants. Institutional review board approval and informed consent has been obtained and patient confidentiality ensured.

Characterising the health and social care segment of the BCS (The Chartered Institute for IT) membership and their continuing professional development needs.

OBJECTIVES: The aim of this study was to identify and characterise the health and social care membership of the British Computer Society (BCS), an international informatics professional organisation, and to determine their ongoing development needs. METHODS: A prepiloted online survey included items on professional regulatory body, job role, work sector, qualifications, career stage, BCS membership (type, specialist group/branch activity (committees, event attendance)), use of BCS.org career planning/continuing professional development (CPD) tools, self-reported digital literacy and other professional registrations. The quantitative data were analysed using descriptive statistics in JASP V.0.9.2 to report frequencies and correlations. RESULTS: Responses were received from 152 participants. Most were male (n=103; 68%), aged 50-59 years (n=41; 28%), working in England (n=107; 71%) with master's or honours degrees (n=80; 53%). Most were either new (5 years or less; n=61; 40%) or long-term members (21 years or more; n=43; 28%) of BCS. Most were not interested in health specialist groups (n=57; 38%) preferring non-health specialist groups such as information management (n=54; 37%) and project management (n=52; 34%). DISCUSSION: This is the first paper to characterise the health and social care membership of an IT-focused professional body and to start to determine their CPD needs. There are further challenges ahead in curating the content and delivery. CONCLUSION: This study is the starting point from which members' CPD needs, and ongoing interest, in being recognised as health and social care professional members, can be acknowledged and explored. Further research is planned with the participants who volunteered to be part of designing future CPD content and delivery.

Machine learning using longitudinal prescription and medical claims for the detection of non-alcoholic steatohepatitis (NASH).

OBJECTIVES: To develop and evaluate machine learning models to detect patients with suspected undiagnosed non-alcoholic steatohepatitis (NASH) for diagnostic screening and clinical management. METHODS: In this retrospective observational non-interventional study using administrative medical claims data from 1 463 089 patients, gradient-boosted decision trees were trained to detect patients with likely NASH from an at-risk patient population with a history of obesity, type 2 diabetes mellitus, metabolic disorder or non-alcoholic fatty liver (NAFL). Models were trained to detect likely NASH in all at-risk patients or in the subset without a prior NAFL diagnosis (at-risk non-NAFL patients). Models were trained and validated using retrospective medical claims data and assessed using area under precision recall curves and receiver operating characteristic curves (AUPRCs and AUROCs). RESULTS: The 6-month incidences of NASH in claims data were 1 per 1437 at-risk patients and 1 per 2127 at-risk non-NAFL patients . The model trained to detect NASH in all at-risk patients had an AUPRC of 0.0107 (95% CI 0.0104 to 0.0110) and an AUROC of 0.84. At 10% recall, model precision was 4.3%, which is 60× above NASH incidence. The model trained to detect NASH in the non-NAFL cohort had an AUPRC of 0.0030 (95% CI 0.0029 to 0.0031) and an AUROC of 0.78. At 10% recall, model precision was 1%, which is 20× above NASH incidence. CONCLUSION: The low incidence of NASH in medical claims data corroborates the pattern of NASH underdiagnosis in clinical practice. Claims-based machine learning could facilitate the detection of patients with probable NASH for diagnostic testing and disease management.

Codeveloping an effective EMPA to maturity in an acute NHS Trust: an implementer report.

INTRODUCTION: University Hospitals Leicester has codeveloped, with Nervecentre, an Electronic Prescribing and Medicines Administration System that meets specific clinical and interoperability demands of the National Health Service (NHS). METHODS: The system was developed through a frontline-led and agile approach with a project team consisting of clinicians, Information Technology (IT) specialists and the vendor's representatives over an 18-month period. RESULTS: The system was deployed successfully with more than a thousand transcriptions during roll-out. Despite the high caseload and novelty of the system, there was no increase in error rates within the first 3 months of roll-out. Healthcare professionals perceived the new system as efficient with improved clinical workflow, and safe through an integrated medication alert system. DISCUSSION: This case study demonstrates how NHS trusts can successfully co-develop, with vendors, new IT systems which meet interoperability standards such as Fast Healthcare Interoperability Resources, while improving front line clinical experience. CONCLUSION: Alternative methods to the 'big bang' deployment of IT projects, such as 'gradual implementation', must be demonstrated and evaluated for their ability to deliver digital transformation projects in the NHS successfully.

Understanding barriers of receiving short message service appointment reminders across African regions: a systematic review.

OBJECTIVE: Patients frequently miss their medical appointments. Therefore, short message service (SMS) has been used as a strategy for medical and healthcare service appointment reminders. This systematic review aimed to identify barriers to SMS appointment reminders across African regions. METHODS: PubMed, Google Scholar, Semantic Scholar and Web of Science were used for searching, and hand searching was done. Original studies written in English, conducted in Africa, and published since 1 December 2018, were included. The standard quality assessment checklist was used for the quality appraisal of the included studies. The Preferred Reporting Items for Systematic Reviews and Meta-Analyses flowchart diagram was used for study selection and screening, and any disagreements were resolved via discussions. RESULTS: A total of 955 articles were searched, 521 studies were removed due to duplication and 105 studies were assessed for eligibility. Consequently, nine studies met the inclusion criteria. Five out of nine included studies were done by randomised control trials. The barriers that hampered patients, mothers and other parental figures of children when they were notified via SMS of medical and health services were identified. Among the 11 identified barriers, illiteracy, issues of confidentiality, familiarised text messages, inadequate information communication technology infrastructure, being a rural resident and loss of mobile phones occurred in at least two studies. CONCLUSIONS: SMS is an effective and widely accepted appointment reminder tool. However, it is hampered by numerous barriers. Hence, we gathered summarised information about users' barriers to SMS-based appointment reminders. Therefore, stakeholders should address existing identified barriers for better Mhealth interventions. PROSPERO REGISTRATION NUMBER: CRD42022296559.

Evaluating algorithmic fairness in the presence of clinical guidelines: the case of atherosclerotic cardiovascular disease risk estimation.

OBJECTIVES: The American College of Cardiology and the American Heart Association guidelines on primary prevention of atherosclerotic cardiovascular disease (ASCVD) recommend using 10-year ASCVD risk estimation models to initiate statin treatment. For guideline-concordant decision-making, risk estimates need to be calibrated. However, existing models are often miscalibrated for race, ethnicity and sex based subgroups. This study evaluates two algorithmic fairness approaches to adjust the risk estimators (group recalibration and equalised odds) for their compatibility with the assumptions underpinning the guidelines' decision rules.MethodsUsing an updated pooled cohorts data set, we derive unconstrained, group-recalibrated and equalised odds-constrained versions of the 10-year ASCVD risk estimators, and compare their calibration at guideline-concordant decision thresholds. RESULTS: We find that, compared with the unconstrained model, group-recalibration improves calibration at one of the relevant thresholds for each group, but exacerbates differences in false positive and false negative rates between groups. An equalised odds constraint, meant to equalise error rates across groups, does so by miscalibrating the model overall and at relevant decision thresholds. DISCUSSION: Hence, because of induced miscalibration, decisions guided by risk estimators learned with an equalised odds fairness constraint are not concordant with existing guidelines. Conversely, recalibrating the model separately for each group can increase guideline compatibility, while increasing intergroup differences in error rates. As such, comparisons of error rates across groups can be misleading when guidelines recommend treating at fixed decision thresholds. CONCLUSION: The illustrated tradeoffs between satisfying a fairness criterion and retaining guideline compatibility underscore the need to evaluate models in the context of downstream interventions.

Investigating for bias in healthcare algorithms: a sex-stratified analysis of supervised machine learning models in liver disease prediction.

OBJECTIVES: The Indian Liver Patient Dataset (ILPD) is used extensively to create algorithms that predict liver disease. Given the existing research describing demographic inequities in liver disease diagnosis and management, these algorithms require scrutiny for potential biases. We address this overlooked issue by investigating ILPD models for sex bias. METHODS: Following our literature review of ILPD papers, the models reported in existing studies are recreated and then interrogated for bias. We define four experiments, training on sex-unbalanced/balanced data, with and without feature selection. We build random forests (RFs), support vector machines (SVMs), Gaussian Naïve Bayes and logistic regression (LR) classifiers, running experiments 100 times, reporting average results with SD. RESULTS: We reproduce published models achieving accuracies of >70% (LR 71.31% (2.37 SD) - SVM 79.40% (2.50 SD)) and demonstrate a previously unobserved performance disparity. Across all classifiers females suffer from a higher false negative rate (FNR). Presently, RF and LR classifiers are reported as the most effective models, yet in our experiments they demonstrate the greatest FNR disparity (RF; -21.02%; LR; -24.07%). DISCUSSION: We demonstrate a sex disparity that exists in published ILPD classifiers. In practice, the higher FNR for females would manifest as increased rates of missed diagnosis for female patients and a consequent lack of appropriate care. Our study demonstrates that evaluating biases in the initial stages of machine learning can provide insights into inequalities in current clinical practice, reveal pathophysiological differences between the male and females, and can mitigate the digitisation of inequalities into algorithmic systems. CONCLUSION: Our findings are important to medical data scientists, clinicians and policy-makers involved in the implementation medical artificial intelligence systems. An awareness of the potential biases of these systems is essential in preventing the digital exacerbation of healthcare inequalities.