Evaluating clinical decision support systems
Distinguish technical performance from clinical usefulness, and examine validation, human factors and monitoring.
A predictive model does not necessarily constitute a clinical intervention, but it can form part of a decision-support system. Other systems follow fixed rules, issue reminders or flag information for attention. Evaluating these tools means asking what happens when people use them in practice.
A strong fit to stored data does not establish that an output will improve a decision. Usability, reliability and integration with existing workflows also matter. The evaluation needs to connect performance with the task for which the system is intended.
Define the role of the system
The description should distinguish the intended users from the target population and explain the inputs, outputs and place in the workflow. A system may provide information, rank possible actions or prompt further assessment. Recording the software and model version makes clear what was evaluated.
The consequences of unavailable, uncertain or incorrect output determine the evidence and oversight needed. A routine reminder and a tool influencing an urgent decision raise different demands.
Input timing is equally important. Data available in a research dataset may not be available when a decision has to be made. Information collected after the target event can produce an optimistic assessment of real-world performance. Understanding the provenance and timing of the inputs is part of evaluating the claim.
Examine development and validation
Development data need an account of how they were collected, labelled and divided. The evaluation dataset must be genuinely separate from data used to train, select or tune the model. Repeated adjustments after inspecting test results compromise the independence of that test.
External validation applies the model to another relevant dataset or setting, investigating how well it generalises beyond the original sample. Performance may depend on the original participants, equipment or local practices. Comparisons across relevant groups can reveal weaknesses hidden by an overall average.
Model performance includes discrimination and calibration. Discrimination refers to the ability to distinguish between people who experienced an event and those who did not. Calibration assesses agreement between estimated probabilities and observed frequencies. A model may order people reasonably while consistently estimating risks too high or too low.
Put performance in context
Consider a fictional alert evaluated across 1,000 encounters. Suppose the target event occurs in 50. The alert identifies 40 of those events but also flags 190 encounters without the event. Its sensitivity is 80%, and 40 of the 230 positive alerts correspond to the event: about 17%.
Identifying a useful threshold depends on the task and the goals of those using the system. False alarms take time to assess, missed events have consequences and a response requires resources. Optimising a statistical measure alone does not establish that the chosen threshold suits the service.
Prevalence affects the interpretation of positive alerts. Performance in a selected research sample may give a misleading impression of the alerts users will see in day-to-day work.
Evaluate people using the system
Models providing actionable information need consideration of usability and integration with existing workflows. Live evaluation can uncover confusing displays, delays, incomplete inputs and workarounds that a retrospective dataset cannot show. Users may also rely too heavily on an output. Training, oversight and the ability to challenge or override a suggestion are part of the intervention.
DECIDE-AI provides reporting guidance for early clinical evaluation of AI decision support. CONSORT-AI addresses reports of randomised trials involving AI interventions. They encourage clear descriptions of systems, input handling, human interaction and evaluation methods. They are reporting guidelines, rather than evidence that a system is safe or effective.
Clinical utility concerns the value a system adds to care. A claim to improve outcomes needs an appropriate comparison with current practice, including patient-relevant outcomes, staff workload and unintended effects. Demonstrating that software can be used in a pilot does not, by itself, establish clinical benefit.
Plan continuing assessment
Ongoing monitoring matters because performance can change across clinical contexts and over time. Changes in populations, records, devices or care pathways can make an earlier evaluation less relevant. Responsibility needs to be clear for reviewing performance, investigating failures, reporting incidents and assessing updates.
A materially changed version needs appropriate evaluation rather than automatically inheriting the evidence for its predecessor. A useful appraisal states the system's functions, where it was tested, its performance and the uncertainties that remain. Claims should be proportionate to the available evidence, including whether use improves relevant decisions and outcomes in the intended setting.
Methodology references
- Vasey et al. DECIDE-AI reporting guideline for early-stage clinical evaluation
- Liu et al. CONSORT-AI extension for clinical trial reports
- NICE: Evidence standards framework for digital health technologies, how to meet the standards