Background
The landscape of clinical decision support systems (CDSSs) is evolving to include increasingly sophisticated data-driven methods, such as machine learning, to provide clinicians with predictions about patients’ risk for negative outcomes or their likely responses to treatments (predictive CDSSs). However, trust in predictive CDSSs has shown to challenge clinician adoption of these tools, precluding the ability to positively impact patient outcomes. This is particularly salient in the hospital setting where clinician time is scarce, and predictive CDSSs have the potential to decrease preventable mortality. Many have advised that clinicians should be involved in the development, implementation, and evaluation of predictive CDSSs to increase translation from development to adoption. Yet, little is known about the prevalence of clinician involvement or the factors that influence clinicians’ trust in predictive CDSSs for the hospital setting. The specific aims of this dissertation were: (a) to survey the literature on predictive CDSSs for the hospital setting to describe the prevalence and methods of clinician involvement throughout stages of system design, (b) to identify and characterize factors that influence clinicians’ trust in predictive CDSSs for in-hospital deterioration, and (c) to explore the use of a trust conceptual framework for incorporating clinician expertise into machine learning model development for predicting rapid response activation among hospitalized non-ICU patients using electronic health record (EHR) data.
Methods
To address the first aim (presented in Chapter Two), a scoping review was conducted to summarize the state of the science of clinician (nurse, physician, physician assistant, nurse practitioner) involvement in predictive CDSS design, with a specific focus on systems using machine learning methods with EHR data for in-hospital decision-making. To address the second aim (presented in Chapter Three), semi-structured interviews with nurses and prescribing providers (i.e., physicians, physicians assistants, nurse practitioners) were conducted and analyzed inductively and deductively (using the Human-Computer Trust conceptual framework) to identify factors that influence trust in predictive CDSSs, using an implemented predictive CDSS for in-hospital deterioration as a grounding example. Finally, to address the third aim (presented in Chapter Four), clinician expertise was elicited in the form of model specifications (requirements, insights, preferences) for facilitating factors shown to influence trust in predictive CDSSs, as guided by the Human-Computer Trust conceptual framework. Specifications included: (a) importance ranking of input features, (b) preference for a more sensitive or specific model, (c) acceptable false positive and negative rates, and (d) prediction lead time. Specifications informed development and evaluation of machine learning models predicting rapid response activation using retrospective EHR data.
Results
The scoping review identified 80 studies. Seventy-six studies described developing a machine learning model for a predictive CDSS, 28% of which described involving clinicians during development. Clinician involvement during development was categorized as: (a) determining clinical relevance/correctness, (b) feature selection, (c) data preprocessing, and (d) serving as a gold standard. Only five studies described implemented predictive CDSSs and no studies described systems in routine use. The qualitative investigation with 17 clinicians (9 prescribing providers, 8 nurses) confirmed that the Human-Computer Trust concepts of perceived understandability and perceived technical competence are factors that influence hospital clinicians’ trust in predictive CDSSs and further characterized these factors (i.e., themes). This study also identified three additional themes influencing trust: (a) actionability, (b) evidence, and (c) equitability, and found that clinicians’ needs for explanations of machine learning models and the impact of discordant predictions may vary according to the extent to which clinicians rely on the predictive CDSS for decision-making. Only two of 28 categories/sub-categories and one theme emerged uniquely to nurses or prescribing providers. Finally, the third study elicited model specifications from fifteen total clinicians. Not all clinicians answered all questions. Vital sign frequency was ranked the most important feature category on average (n = 8 clinicians), the most frequently preferred prediction lead time was shift-change/8-12 hours (n = 9 clinicians), most preferred a more specific than sensitive model (71%; n = 7 clinicians), the average acceptable false positive rate was 42% (n = 9 clinicians), the average acceptable false negative rate was 29% (n = 6 clinicians). These specifications informed development and testing of four machine learning classification models (ridge regression, decision trees, random forest, and XGBoost). 249,676 patient admissions from 2015–2018 at a large northeastern hospital system were modeled to predict whether or not patients would have a rapid response within the 12-hour shift. The random forest classifier met clinician’s average acceptable false positive (27.7%) and negative rates (28.9%) and was marginally more specific (72.2%) than sensitive (71.1%) on a holdout test set.
Conclusions
Studies do not routinely report clinician involvement in model development of predictive CDSSs for the hospital setting and publications on implementation considerably lag those on development. Nurses and prescribing providers described largely shared experiences of trust in predictive CDSSs. Clinicians’ reliance on the predictive CDSS for decision-making within the target clinical workflow should be considered when aiming to facilitate trust. Incorporating clinician expertise into model development for the purpose of facilitating trust is feasible. Future research is needed on the impact of clinician involvement on trust, clinicians’ personal attributes that influence trust, and explanation design. Increased education for clinicians about predictive CDSSs is recommended.
| Identifer | oai:union.ndltd.org:columbia.edu/oai:academiccommons.columbia.edu:10.7916/d8-tffc-br50 |
| Date | January 2021 |
| Creators | Schwartz, Jessica |
| Source Sets | Columbia University |
| Language | English |
| Detected Language | English |
| Type | Theses |
Page generated in 0.0027 seconds