Table 5.

ML-based clinical tool evaluation framework steps.

Step 1: Define clinical quality improvement goal and opportunity unlocked by predicting the clinical event
Considerations	Impact
• Define patient care/quality goals • Identify actionable clinical events that if predicted help achieve goals • Establish metrics and results required to identify “at risk” patients • Evaluate if this type of tool is useful for furthering goals • Determine how the tool will embed into clinical workflow, and what actions need to be taken based on predicted clinical event • Define key metrics for evaluating clinical impact of risk predictions	All stakeholders (clinicians, business leaders, data scientists, etc.,) will have a clear understanding of how deployment of an ML-based clinical tool will help to achieve quality improvement goals. Teams should be able to fill in this statement: “If the care team knows that X event will happen, they will take Y action, to increase Z value”.
Step 2: Build/acquire ML-based clinical tool that predicts defined clinical event
Considerations	Impact
Decide whether to build a custom ML-based tool or acquire an existing ML-based tool that is practical, customizable, and suited for the practice’s local data patterns	Organization will be equipped with the right ML-based clinical tool for their intended goals
Step 3: Conduct retrospective evaluation of ML-based clinical tool
Considerations	Impact
Retrospectively apply model to a representative historical patient population from the institution and then compare predictions with known past observed events to confirm if the tool meets desired metrics	Allows the organization to expediently assess the suitability of the ML-based clinical tool for the prediction task at hand
Step 4: Conduct bias assessment
Considerations	Impact
• Proactively evaluate for bias, including treatment pattern disparities or lack of representation, choice of modeling approach, or choice of predicted clinical event • Make necessary adjustments to the tool before there is any impact on patients	Ensures that the ML-based clinical tool algorithms do not reproduce real-world inequalities that can occur as a result of treatment pattern disparities or a lack of representation encoded in datasets, the choice of modeling approach, or the choice of predicted clinical event
Step 5: Conduct prospective evaluation of ML-based tool
Considerations	Impact
• Conduct a prospective evaluation on a present-day, real-world patient population in a randomized setting to understand how well the model is likely to perform in real time • Note: This step may not be necessary in every case if the ML-based tool has been prospectively evaluated and its performance in real-world setting monitored	Prospective validation is considered the “gold standard” of ML model validation when applied to the point-of-care setting because it shows the clinical team how well the model is likely to perform in real time where several factors can affect model performance, such as recent pattern changes in the real world (e.g., occurrence of a pandemic), care delivery (e.g., updates to clinical standards), or technical or operational issues (e.g., data entry delays that can make a system unusable in practice)
Step 6: Embed and monitor tool in clinical workflow
Considerations	Impact
• Adopt tool into standard clinical workflow • Conduct data quality monitoring, performance monitoring, and bias monitoring • The ML-based tool should not replace traditional patient identification processes, but support them with a data-driven approach that also enhances their efficiency	The ML-based tool can now be used to achieve the quality improvement goal defined in Step 1. Ongoing monitoring ensures the model’s suitability in the dynamic clinical environment of the real world where patterns of care seeking and care delivery evolve, and that model predictions are not impacted by manual or technical errors that could inadvertently affect a patient’s predicted risk and/or access to supplemental care.

ML machine learning.