From: Use of artificial intelligence in critical care: opportunities and obstacles
Phase of development | Obstacle | Cause | Work around | Aspirational solution |
---|---|---|---|---|
Planning | Identify ideal source population | Too narrow a target population Inappropriate selection of outcome measures | Involve a diverse group of stakeholders with different perspectives to identify shortcomings and goals | Involve diverse stakeholders to plan data collection |
Inherent bias in data | Limited number of variables included in the data source Limited availability of relevant data elements (e.g., Social Determinants of Health) Poor quality of data dictionary | Produce data cards Identify non-traditional sources of data Improve data dictionaries | Ongoing construction of large, diverse, bias-aware, multimodal datasets | |
Data source has limited diversity and marginally relevant to target population | Limited number of large databases available Poor choice of source data | Identify additional data sources Propensity matching | ||
Variability in source data content and formatting | Homegrown, narrowly focused data dictionaries and case report forms No concern for reuse | Post-hoc harmonization of data | Adoption of a standard common data model for harmonization | |
Data | Data source identification | Existing data not appropriate for research question Data is difficult to find | Use and modify data that can be found | Large data distribution platforms |
Data access | Unclear governance Data may be impossible to access | Data use agreements Document investigators’ qualifications, giving tiered access to data | Transparent and uniform governance | |
Privacy | Heterogeneity of privacy rules Technical issues with data sharing Lack of guidelines | Awareness of regional privacy requirements Adoption of sufficient technical solution to fulfill these requirements Contract technical expertise | Democratized low-cost technical solutions | |
Data quality | Poor planning Poor collection methodology | Data imputation Limit inference | Adoption of existing tools Adoption of guidelines on data quality assessment Develop guidelines for emerging domains of data | |
Data Reuse | Limited use of common data models | Post-hoc harmonization Promote easy access | Abide by FAIR principles (www.go-fair.org) | |
Model development | Technical performance | Choice of performance metric inappropriate for intended CDSS use Risk of over-fitting and lack of generalizability | Model cards Follow existing guidelines for reporting AI model performance | Good AI methodology stewardship |
Algorithmic bias | Inherent or pre-processing bias in training dataset Possible bias not assessed Tools for bias assessment may be incomplete | Use available tools to assess bias at different stages Tune models on subgroups on interest | Guidelines on multi-stage bias assessment Disseminate known limits, shortcomings of CDSS | |
Trust and Explainability | Lack of trust in black box models Poor performance of explainable models Output not seen as useful or valuable by end-users | Expanding sets of methods to explain models Chose explainable models whenever decrease in performance not clinically relevant Engage end-users very early in the CDSS development | Team development from initiation to completion of model development | |
Prospective evaluation | Technical access to relevant real-time data Governance barriers to real-time data Latency in accessing real-time data | Simulate real-time data arrival Silent deployment | Real-time EHR integration for deployment in a learning healthcare system | |
CDSS | Real-time benchmarking | Technical obstacles to deliver model prediction to bedside Governance-based obstacles to bedside deployment Ethical issues with real-time deployment | Early involvement of key administrative, technical, and clinical stakeholders | Flexible pipeline for scalable expansion of tools Improved technical standards for real-time integration of AI applications with EHR vendors Wide adoption of such standards |
Transitioning a model into a CDSS | Human factor engineering Implementation | Involvement of end-users in system design Choosing appropriate metrics of performance System usability and technology acceptance evaluations | Systematic usability and acceptability assessments Periodic reassessments | |
Effectiveness assessment | Measuring impact | Impact on clinical outcomes Impact on ICU processes Impact on technical team | Choosing appropriate clinical outcomes Measure resource use Choosing appropriate technical outcomes | Engage stakeholder to define a process for technology effectiveness assessments National guidance for technology effectiveness assessment Ongoing compliance assessment |
Adoption | Compliance | Identify internal “super users” to lead adoption Involve end-users in evaluation | ||
Unclear return on investment | Impact of economic outcomes | Engage financial decision makers early in system design Collect data that allows economic assessment | ||
Real-world deployment | CDSS degrading performance over time | Patient-specific issues changing or care giver support decreasing | Review model performance and address changes in input variables | Continual CDSS model performance evaluation |
Modify CDSS to be more ergonomic for caregivers Re-educate caregivers | ||||
Commercial availability | Many solutions are homegrown Implementation gap and lack of practical solutions FDA, GDPR clearance | Clinicians’ involvement in product development and testing AI-technology assessment bodies in Health system that prioritize development | Evolving national guidance on product development and commercial clearance Inherent cyber security of EHR-linked algorithms used in modeling | |
All phases | Limited technical expertise | Inherent lack of expertise of clinicians in big data manipulation and model development Siloed development | Build interdisciplinary teams Broadly available educational opportunities | Basic training in AI in clinical curricula Early engagement of potential stakeholders Worldwide datathons |
Limited scalability | Solution not applicable outside ICUs or in resource limited environments | Conceive of critical illness as a continuum and formulate outcomes accordingly Prioritize low-technology solutions |