A framework for the oversight and local deployment of safe and high-quality prediction models

Transparent medical image AI via an image-text foundation model grounded in medical literature

Building trustworthy and transparent image-based medical artificial intelligence (AI) systems requires the ability to interrogate data and models at all stages of the development pipeline, from training models to post-deployment monitoring. Ideally, the data and associated AI systems could be described using terms already familiar to physicians, but this requires medical datasets densely annotated with semantically meaningful concepts. In the present study, we present a foundation model approach, named MONET (medical concept retriever), which learns how to connect medical images with text and densely scores images on concept presence to enable important tasks in medical AI development and deployment such as data auditing, model auditing and model interpretation. Dermatology provides a demanding use case for the versatility of MONET, due to the heterogeneity in diseases, skin tones and imaging modalities. We trained MONET based on 105,550 dermatological images paired with natural language descriptions from a large collection of medical literature. MONET can accurately annotate concepts across dermatology images as verified by board-certified dermatologists, competitively with supervised models built on previously concept-annotated dermatology datasets of clinical images. We demonstrate how MONET enables AI transparency across the entire AI system development pipeline, from building inherently interpretable models to dataset and model auditing, including a case study dissecting the results of an AI clinical trial.

Interpretable Predictive Models to Understand Risk Factors for Maternal and Fetal Outcomes

Although most pregnancies result in a good outcome, complications are not uncommon and can be associated with serious implications for mothers and babies. Predictive modeling has the potential to improve outcomes through a better understanding of risk factors, heightened surveillance for high-risk patients, and more timely and appropriate interventions, thereby helping obstetricians deliver better care. We identify and study the most important risk factors for four types of pregnancy complications: (i) severe maternal morbidity, (ii) shoulder dystocia, (iii) preterm preeclampsia, and (iv) antepartum stillbirth. We use an Explainable Boosting Machine (EBM), a high-accuracy glass-box learning method, for the prediction and identification of important risk factors. We undertake external validation and perform an extensive robustness analysis of the EBM models. EBMs match the accuracy of other black-box ML methods, such as deep neural networks and random forests, and outperform logistic regression, while being more interpretable. EBMs prove to be robust. The interpretability of the EBM models reveal surprising insights into the features contributing to risk (e.g., maternal height is the second most important feature for shoulder dystocia) and may have potential for clinical application in the prediction and prevention of serious complications in pregnancy.

Empirical data drift detection experiments on real-world medical imaging data

While it is common to monitor deployed clinical artificial intelligence (AI) models for performance degradation, it is less common for the input data to be monitored for data drift - systemic changes to input distributions. However, when real-time evaluation may not be practical (eg., labeling costs) or when gold-labels are automatically generated, we argue that tracking data drift becomes a vital addition for AI deployments. In this work, we perform empirical experiments on real-world medical imaging to evaluate three data drift detection methods' ability to detect data drift caused (a) naturally (emergence of COVID-19 in X-rays) and (b) synthetically. We find that monitoring performance alone is not a good proxy for detecting data drift and that drift-detection heavily depends on sample size and patient features. Our work discusses the need and utility of data drift detection in various scenarios and highlights gaps in knowledge for the practical application of existing methods.

Designing an Implementable Clinical Prediction Model for Near-Term Mortality and Long-Term Survival in Patients on Maintenance Hemodialysis

RATIONALE & OBJECTIVE

The life expectancy of patients treated with maintenance hemodialysis (MHD) is heterogeneous. Knowledge of life-expectancy may focus care decisions on near-term versus long-term goals. The current tools are limited and focus on near-term mortality. Here, we develop and assess potential utility for predicting near-term mortality and long-term survival on MHD.

STUDY DESIGN

Predictive modeling study.

SETTING & PARTICIPANTS

42,351 patients contributing 997,381 patient months over 11 years, abstracted from the electronic health record (EHR) system of midsize, nonprofit dialysis providers.

NEW PREDICTORS & ESTABLISHED PREDICTORS

Demographics, laboratory results, vital signs, and service utilization data available within dialysis EHR.

OUTCOME

For each patient month, we ascertained death within the next 6 months (ie, near-term mortality) and survival over more than 5 years during receipt of MHD or after kidney transplantation (ie, long-term survival).

ANALYTICAL APPROACH

We used least absolute shrinkage and selection operator logistic regression and gradient-boosting machines to predict each outcome. We compared these to time-to-event models spanning both time horizons. We explored the performance of decision rules at different cut points.

RESULTS

All models achieved an area under the receiver operator characteristic curve of≥0.80 and optimal calibration metrics in the test set. The long-term survival models had significantly better performance than the near-term mortality models. The time-to-event models performed similarly to binary models. Applying different cut points spanning from the 1st to 90th percentile of the predictions, a positive predictive value (PPV) of 54% could be achieved for near-term mortality, but with poor sensitivity of 6%. A PPV of 71% could be achieved for long-term survival with a sensitivity of 67%.

LIMITATIONS

The retrospective models would need to be prospectively validated before they could be appropriately used as clinical decision aids.

CONCLUSIONS

A model built with readily available clinical variables to support easy implementation can predict clinically important life expectancy thresholds and shows promise as a clinical decision support tool for patients on MHD. Predicting long-term survival has better decision rule performance than predicting near-term mortality.

PLAIN-LANGUAGE SUMMARY

Clinical prediction models (CPMs) are not widely used for patients undergoing maintenance hemodialysis (MHD). Although a variety of CPMs have been reported in the literature, many of these were not well-designed to be easily implementable. We consider the performance of an implementable CPM for both near-term mortality and long-term survival for patients undergoing MHD. Both near-term and long-term models have similar predictive performance, but the long-term models have greater clinical utility. We further consider how the differential performance of predicting over different time horizons may be used to impact clinical decision making. Although predictive modeling is not regularly used for MHD patients, such tools may help promote individualized care planning and foster shared decision making.

Translating ethical and quality principles for the effective, safe and fair development, deployment and use of artificial intelligence technologies in healthcare

OBJECTIVE

The complexity and rapid pace of development of algorithmic technologies pose challenges for their regulation and oversight in healthcare settings. We sought to improve our institution's approach to evaluation and governance of algorithmic technologies used in clinical care and operations by creating an Implementation Guide that standardizes evaluation criteria so that local oversight is performed in an objective fashion.

MATERIALS AND METHODS

Building on a framework that applies key ethical and quality principles (clinical value and safety, fairness and equity, usability and adoption, transparency and accountability, and regulatory compliance), we created concrete guidelines for evaluating algorithmic technologies at our institution.

RESULTS

An Implementation Guide articulates evaluation criteria used during review of algorithmic technologies and details what evidence supports the implementation of ethical and quality principles for trustworthy health AI. Application of the processes described in the Implementation Guide can lead to algorithms that are safer as well as more effective, fair, and equitable upon implementation, as illustrated through 4 examples of technologies at different phases of the algorithmic lifecycle that underwent evaluation at our academic medical center.

DISCUSSION

By providing clear descriptions/definitions of evaluation criteria and embedding them within standardized processes, we streamlined oversight processes and educated communities using and developing algorithmic technologies within our institution.

CONCLUSIONS

We developed a scalable, adaptable framework for translating principles into evaluation criteria and specific requirements that support trustworthy implementation of algorithmic technologies in patient care and healthcare operations.

President Biden's Executive Order on Artificial Intelligence-Implications for Health Care Organizations

This Viewpoint discusses a recent executive order by US President Joe Biden about the development and implementation of AI, including the role of government vs the private sector and how the order may affect health care.

Dashboarding to Monitor Machine-Learning-Based Clinical Decision Support Interventions

BACKGROUND

 Existing monitoring of machine-learning-based clinical decision support (ML-CDS) is focused predominantly on the ML outputs and accuracy thereof. Improving patient care requires not only accurate algorithms but also systems of care that enable the output of these algorithms to drive specific actions by care teams, necessitating expanding their monitoring.

OBJECTIVES

 In this case report, we describe the creation of a dashboard that allows the intervention development team and operational stakeholders to govern and identify potential issues that may require corrective action by bridging the monitoring gap between model outputs and patient outcomes.

METHODS

 We used an iterative development process to build a dashboard to monitor the performance of our intervention in the broader context of the care system.

RESULTS

 Our investigation of best practices elsewhere, iterative design, and expert consultation led us to anchor our dashboard on alluvial charts and control charts. Both the development process and the dashboard itself illuminated areas to improve the broader intervention.

CONCLUSION

 We propose that monitoring ML-CDS algorithms with regular dashboards that allow both a context-level view of the system and a drilled down view of specific components is a critical part of implementing these algorithms to ensure that these tools function appropriately within the broader care system.

How patients distinguish between clinical and administrative predictive models in health care

OBJECTIVES

To understand patient perceptions of specific applications of predictive models in health care.

STUDY DESIGN

Original, cross-sectional national survey.

METHODS

We conducted a national online survey of US adults with the National Opinion Research Center from November to December 2021. Measures of internal consistency were used to identify how patients differentiate between clinical and administrative predictive models. Multivariable logistic regressions were used to identify relationships between comfort with various types of predictive models and patient demographics, perceptions of privacy protections, and experiences in the health care system.

RESULTS

A total of 1541 respondents completed the survey. After excluding observations with missing data for the variables of interest, the final analytic sample was 1488. We found that patients differentiate between clinical and administrative predictive models. Comfort with prediction of bill payment and missed appointments was especially low (21.6% and 36.6%, respectively). Comfort was higher with clinical predictive models, such as predicting stroke in an emergency (55.8%). Experiences of discrimination were significant negative predictors of comfort with administrative predictive models. Health system transparency around privacy policies was a significant positive predictor of comfort with both clinical and administrative predictive models.

CONCLUSIONS

Patients are more comfortable with clinical applications of predictive models than administrative ones. Privacy protections and transparency about how health care systems protect patient data may facilitate patient comfort with these technologies. However, larger inequities and negative experiences in health care remain important for how patients perceive administrative applications of prediction.

Guiding Principles to Address the Impact of Algorithm Bias on Racial and Ethnic Disparities in Health and Health Care

Importance

Health care algorithms are used for diagnosis, treatment, prognosis, risk stratification, and allocation of resources. Bias in the development and use of algorithms can lead to worse outcomes for racial and ethnic minoritized groups and other historically marginalized populations such as individuals with lower income.

Objective

To provide a conceptual framework and guiding principles for mitigating and preventing bias in health care algorithms to promote health and health care equity.

Evidence Review

The Agency for Healthcare Research and Quality and the National Institute for Minority Health and Health Disparities convened a diverse panel of experts to review evidence, hear from stakeholders, and receive community feedback.

Findings

The panel developed a conceptual framework to apply guiding principles across an algorithm's life cycle, centering health and health care equity for patients and communities as the goal, within the wider context of structural racism and discrimination. Multiple stakeholders can mitigate and prevent bias at each phase of the algorithm life cycle, including problem formulation (phase 1); data selection, assessment, and management (phase 2); algorithm development, training, and validation (phase 3); deployment and integration of algorithms in intended settings (phase 4); and algorithm monitoring, maintenance, updating, or deimplementation (phase 5). Five principles should guide these efforts: (1) promote health and health care equity during all phases of the health care algorithm life cycle; (2) ensure health care algorithms and their use are transparent and explainable; (3) authentically engage patients and communities during all phases of the health care algorithm life cycle and earn trustworthiness; (4) explicitly identify health care algorithmic fairness issues and trade-offs; and (5) establish accountability for equity and fairness in outcomes from health care algorithms.

Conclusions and Relevance

Multiple stakeholders must partner to create systems, processes, regulations, incentives, standards, and policies to mitigate and prevent algorithmic bias. Reforms should implement guiding principles that support promotion of health and health care equity in all phases of the algorithm life cycle as well as transparency and explainability, authentic community engagement and ethical partnerships, explicit identification of fairness issues and trade-offs, and accountability for equity and fairness.

Development, Deployment, and Implementation of a Machine Learning Surgical Case Length Prediction Model and Prospective Evaluation

OBJECTIVE

To implement a machine learning model using only the restricted data available at case creation time to predict surgical case length for multiple services at different locations.

BACKGROUND

The operating room is one of the most expensive resources in a health system, estimated to cost $22 to $133 per minute and generate about 40% of hospital revenue. Accurate prediction of surgical case length is necessary for efficient scheduling and cost-effective utilization of the operating room and other resources.

METHODS

We introduced a similarity cascade to capture the complexity of cases and surgeon influence on the case length and incorporated that into a gradient-boosting machine learning model. The model loss function was customized to improve the balance between over- and under-prediction of the case length. A production pipeline was created to seamlessly deploy and implement the model across our institution.

RESULTS

The prospective analysis showed that the model output was gradually adopted by the schedulers and outperformed the scheduler-predicted case length from August to December 2022. In 33,815 surgical cases across outpatient and inpatient platforms, the operational implementation predicted 11.2% fewer underpredicted cases and 5.9% more cases within 20% of the actual case length compared with the schedulers and only overpredicted 5.3% more. The model assisted schedulers to predict 3.4% more cases within 20% of the actual case length and 4.3% fewer underpredicted cases.

CONCLUSIONS

We created a unique framework that is being leveraged every day to predict surgical case length more accurately at case posting time and could be potentially utilized to deploy future machine learning models.

Artificial Intelligence Governance and Otolaryngology-Head and Neck Surgery

This rapid communication highlights components of artificial intelligence governance in healthcare and suggests adopting key governance approaches in otolaryngology – head and neck surgery.

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.

DEPLOYR: a technical framework for deploying custom real-time machine learning models into the electronic medical record

OBJECTIVE

Heatlhcare institutions are establishing frameworks to govern and promote the implementation of accurate, actionable, and reliable machine learning models that integrate with clinical workflow. Such governance frameworks require an accompanying technical framework to deploy models in a resource efficient, safe and high-quality manner. Here we present DEPLOYR, a technical framework for enabling real-time deployment and monitoring of researcher-created models into a widely used electronic medical record system.

MATERIALS AND METHODS

We discuss core functionality and design decisions, including mechanisms to trigger inference based on actions within electronic medical record software, modules that collect real-time data to make inferences, mechanisms that close-the-loop by displaying inferences back to end-users within their workflow, monitoring modules that track performance of deployed models over time, silent deployment capabilities, and mechanisms to prospectively evaluate a deployed model's impact.

RESULTS

We demonstrate the use of DEPLOYR by silently deploying and prospectively evaluating 12 machine learning models trained using electronic medical record data that predict laboratory diagnostic results, triggered by clinician button-clicks in Stanford Health Care's electronic medical record.

DISCUSSION

Our study highlights the need and feasibility for such silent deployment, because prospectively measured performance varies from retrospective estimates. When possible, we recommend using prospectively estimated performance measures during silent trials to make final go decisions for model deployment.

CONCLUSION

Machine learning applications in healthcare are extensively researched, but successful translations to the bedside are rare. By describing DEPLOYR, we aim to inform machine learning deployment best practices and help bridge the model implementation gap.

Implementation frameworks for end-to-end clinical AI: derivation of the SALIENT framework

OBJECTIVE

To derive a comprehensive implementation framework for clinical AI models within hospitals informed by existing AI frameworks and integrated with reporting standards for clinical AI research.

MATERIALS AND METHODS

(1) Derive a provisional implementation framework based on the taxonomy of Stead et al and integrated with current reporting standards for AI research: TRIPOD, DECIDE-AI, CONSORT-AI. (2) Undertake a scoping review of published clinical AI implementation frameworks and identify key themes and stages. (3) Perform a gap analysis and refine the framework by incorporating missing items.

RESULTS

The provisional AI implementation framework, called SALIENT, was mapped to 5 stages common to both the taxonomy and the reporting standards. A scoping review retrieved 20 studies and 247 themes, stages, and subelements were identified. A gap analysis identified 5 new cross-stage themes and 16 new tasks. The final framework comprised 5 stages, 7 elements, and 4 components, including the AI system, data pipeline, human-computer interface, and clinical workflow.

DISCUSSION

This pragmatic framework resolves gaps in existing stage- and theme-based clinical AI implementation guidance by comprehensively addressing the what (components), when (stages), and how (tasks) of AI implementation, as well as the who (organization) and why (policy domains). By integrating research reporting standards into SALIENT, the framework is grounded in rigorous evaluation methodologies. The framework requires validation as being applicable to real-world studies of deployed AI models.

CONCLUSIONS

A novel end-to-end framework has been developed for implementing AI within hospital clinical practice that builds on previous AI implementation frameworks and research reporting standards.

Effectiveness of an Emergency Department-Based Machine Learning Clinical Decision Support Tool to Prevent Outpatient Falls Among Older Adults: Protocol for a Quasi-Experimental Study

BACKGROUND

Emergency department (ED) providers are important collaborators in preventing falls for older adults because they are often the first health care providers to see a patient after a fall and because at-home falls are often preceded by previous ED visits. Previous work has shown that ED referrals to falls interventions can reduce the risk of an at-home fall by 38%. Screening patients at risk for a fall can be time-consuming and difficult to implement in the ED setting. Machine learning (ML) and clinical decision support (CDS) offer the potential of automating the screening process. However, it remains unclear whether automation of screening and referrals can reduce the risk of future falls among older patients.

OBJECTIVE

The goal of this paper is to describe a research protocol for evaluating the effectiveness of an automated screening and referral intervention. These findings will inform ongoing discussions about the use of ML and artificial intelligence to augment medical decision-making.

METHODS

To assess the effectiveness of our program for patients receiving the falls risk intervention, our primary analysis will be to obtain referral completion rates at 3 different EDs. We will use a quasi-experimental design known as a sharp regression discontinuity with regard to intent-to-treat, since the intervention is administered to patients whose risk score falls above a threshold. A conditional logistic regression model will be built to describe 6-month fall risk at each site as a function of the intervention, patient demographics, and risk score. The odds ratio of a return visit for a fall and the 95% CI will be estimated by comparing those identified as high risk by the ML-based CDS (ML-CDS) and those who were not but had a similar risk profile.

RESULTS

The ML-CDS tool under study has been implemented at 2 of the 3 EDs in our study. As of April 2023, a total of 1326 patient encounters have been flagged for providers, and 339 unique patients have been referred to the mobility and falls clinic. To date, 15% (45/339) of patients have scheduled an appointment with the clinic.

CONCLUSIONS

This study seeks to quantify the impact of an ML-CDS intervention on patient behavior and outcomes. Our end-to-end data set allows for a more meaningful analysis of patient outcomes than other studies focused on interim outcomes, and our multisite implementation plan will demonstrate applicability to a broad population and the possibility to adapt the intervention to other EDs and achieve similar results. Our statistical methodology, regression discontinuity design, allows for causal inference from observational data and a staggered implementation strategy allows for the identification of secular trends that could affect causal associations and allow mitigation as necessary.

TRIAL REGISTRATION

ClinicalTrials.gov NCT05810064; https://www.clinicaltrials.gov/study/NCT05810064.

INTERNATIONAL REGISTERED REPORT IDENTIFIER (IRRID)

DERR1-10.2196/48128.

The Role of Human and Organizational Factors in the Pursuit of One Digital Health

OBJECTIVE

This paper surveys a subset of the 2022 human and organizational factor (HOF) literature to provide guidance on building a One Digital Health ecosystem.

METHODS

We searched a subset of journals in PubMed/Medline for studies with "human factors" or "organization" in the title or abstract. Papers published in 2022 were eligible for inclusion in the survey. Selected papers were categorized into structural and behavioural aspects to understand digital health enabled interactions across micro, meso, and macro systems.

RESULTS

Our survey of the 2022 HOF literature showed that while we continue to make meaningful progress at digital health enabled interactions across systems levels, there are still challenges that must be overcome. For example, we must continue to grow the breadth of HOF research beyond individual users and systems to assist with the scale up of digital health systems across and beyond organizations. We summarize the findings by providing five HOF considerations to help build a One Digital Health ecosystem.

CONCLUSION

One Digital Health challenges us to improve coordination, communication, and collaboration between the health, environmental and veterinary sectors. Doing so requires us to develop both the structural and behavioural capacity of digital health systems at the organizational level and beyond so that we can develop more robust and integrated systems across health, environmental and veterinary sectors. The HOF community has much to offer and must play a leading role in designing a One Digital Health ecosystem.

The Role of Artificial Intelligence Model Documentation in Translational Science: Scoping Review

BACKGROUND

Despite the touted potential of artificial intelligence (AI) and machine learning (ML) to revolutionize health care, clinical decision support tools, herein referred to as medical modeling software (MMS), have yet to realize the anticipated benefits. One proposed obstacle is the acknowledged gaps in AI translation. These gaps stem partly from the fragmentation of processes and resources to support MMS transparent documentation. Consequently, the absence of transparent reporting hinders the provision of evidence to support the implementation of MMS in clinical practice, thereby serving as a substantial barrier to the successful translation of software from research settings to clinical practice.

OBJECTIVE

This study aimed to scope the current landscape of AI- and ML-based MMS documentation practices and elucidate the function of documentation in facilitating the translation of ethical and explainable MMS into clinical workflows.

METHODS

A scoping review was conducted in accordance with PRISMA-ScR (Preferred Reporting Items for Systematic Reviews and Meta-Analyses extension for Scoping Reviews) guidelines. PubMed was searched using Medical Subject Headings key concepts of AI, ML, ethical considerations, and explainability to identify publications detailing AI- and ML-based MMS documentation, in addition to snowball sampling of selected reference lists. To include the possibility of implicit documentation practices not explicitly labeled as such, we did not use documentation as a key concept but as an inclusion criterion. A 2-stage screening process (title and abstract screening and full-text review) was conducted by 1 author. A data extraction template was used to record publication-related information; barriers to developing ethical and explainable MMS; available standards, regulations, frameworks, or governance strategies related to documentation; and recommendations for documentation for papers that met the inclusion criteria.

RESULTS

Of the 115 papers retrieved, 21 (18.3%) papers met the requirements for inclusion. Ethics and explainability were investigated in the context of AI- and ML-based MMS documentation and translation. Data detailing the current state and challenges and recommendations for future studies were synthesized. Notable themes defining the current state and challenges that required thorough review included bias, accountability, governance, and explainability. Recommendations identified in the literature to address present barriers call for a proactive evaluation of MMS, multidisciplinary collaboration, adherence to investigation and validation protocols, transparency and traceability requirements, and guiding standards and frameworks that enhance documentation efforts and support the translation of AI- and ML-based MMS.

CONCLUSIONS

Resolving barriers to translation is critical for MMS to deliver on expectations, including those barriers identified in this scoping review related to bias, accountability, governance, and explainability. Our findings suggest that transparent strategic documentation, aligning translational science and regulatory science, will support the translation of MMS by coordinating communication and reporting and reducing translational barriers, thereby furthering the adoption of MMS.

APLUS: A Python library for usefulness simulations of machine learning models in healthcare

Despite the creation of thousands of machine learning (ML) models, the promise of improving patient care with ML remains largely unrealized. Adoption into clinical practice is lagging, in large part due to disconnects between how ML practitioners evaluate models and what is required for their successful integration into care delivery. Models are just one component of care delivery workflows whose constraints determine clinicians' abilities to act on models' outputs. However, methods to evaluate the usefulness of models in the context of their corresponding workflows are currently limited. To bridge this gap we developed APLUS, a reusable framework for quantitatively assessing via simulation the utility gained from integrating a model into a clinical workflow. We describe the APLUS simulation engine and workflow specification language, and apply it to evaluate a novel ML-based screening pathway for detecting peripheral artery disease at Stanford Health Care.