Using Machine Learning-based Clinical Decision Support for Type 2 Diabetes in Medicaid
How a CDS system resulted in better treatment selection and health outcomes in a Medicaid population
It is snowy day here in Washington, D.C. with 6 inches of snow on the ground and the potential for another six. Welcome back to the Health Tech Happy Hour publication in 2025. We will continue to cover emerging clinical technologies, health policy, and care delivery issues and research this year!
Clinical decision support systems (CDSS) represent one of health care's most promising technological advances, offering the potential to improve patient care by providing clinicians with evidence-based recommendations at the point of decision-making. Recommendations are generated from rules-based logic created from systematic literature reviews, from machine learning analysis of patient data, or from artificial intelligence and deep learning methodologies.
These systems, which range from simple medication alerts to sophisticated treatment selection algorithms (e.g., precision medicine), are designed to enhance clinical decision-making by combining medical knowledge with patient-specific data. However, despite their theoretical promise, the journey from technological innovation to demonstrated clinical benefit and implementation has been remarkably uneven.
This article focuses on the use of machine learning clinical decision support systems in real clinical settings, but if you are interested in LLM technology use in the inpatient and nursing education settings, please see a previous article below:
The Challenge of Prospective Clinical Validation
A significant challenge in the field of health care technology is the frequent disconnect between technical capabilities, pre-clinical validation, and proven clinical outcomes. While many CDSS and machine learning-based solutions demonstrate impressive performance in retrospective analyses, relatively few undergo rigorous clinical validation in real-world health care environments. This validation gap is particularly concerning given the critical nature of medical decision-making and its direct impact on patient outcomes.
The reasons for this validation gap are multifaceted. First, conducting proper clinical validation studies requires significant time, resources, IRB-approval, training, and coordination across multiple stakeholders (i.e., very expensive). Second, the complexity of health care delivery makes it challenging to isolate the specific impact of CDSS interventions from other factors affecting patient care. Finally, there's often pressure to deploy promising technologies quickly, sometimes at the expense of thorough clinical validation.
However, there is an ongoing debate around whether clinical validation of AI and ML CDS technologies should rely on prospective real-world evidence or on tightly controlled randomized clinical trials. One argument for either prospective study design is that prospective, in-clinic studies garner trust to facilitate adoption of CDSS systems that demonstrate effectiveness and usefulness.
Usefulness is important with CDSS as they interface with human clinicians as opposed to drugs where the product itself may only interface with the patient through ingestion and the clinician may witness solely the patient physiological changes.
A Study of CDSS in a Real Clinical Setting for Type 2 Diabetes
A recent study published in the International Journal of Medical Informatics provides a valuable example of how CDSS can be rigorously evaluated in a real clinical setting and can demonstrate positive effects. The research, conducted across multiple health care locations, examined the impact of a CDSS combined with team-based care for managing type 2 diabetes among Medicaid patients.
The study addressed a well-documented challenge in diabetes care: clinical inertia, or the failure to intensify treatment when indicated. Despite clear evidence-based guidelines for diabetes management, studies have shown that patients often experience delays of several years before receiving appropriate treatment intensification when they do not reach key milestones. This problem is particularly acute among Medicaid populations, where approximately 30% of patients with HbA1c levels above 8% experience clinical inertia in their care.
The CDSS and its Use
The CDSS is a comprehensive treatment recommendation system that analyzes over 160 diabetes treatments, capable of suggesting combinations of up to 5 therapies, including diet and exercise. Built on American Association of Clinical Endocrinologists (AACE) and American Diabetes Association (ADA) guidelines, the system generates recommendations based on patient clinical characteristics, preferences, and values, while prioritizing factors like HbA1c control, cost, weight change, adherence, and side effects. In the study, pharmacists primarily used the system in consultation with patients, providing recommendations to physicians who then had four possible response paths: fully following the recommended regimen, partially following it, maintaining the existing regimen, or implementing an alternative treatment plan not suggested by the CDSS.
The CDSS was designed to optimize workflow and support shared decision-making. It considered multiple factors including:
Patient clinical characteristics
Cost sensitivity and budget constraints
Adherence patterns and barriers
Patient preferences and values
Insurance coverage and formulary requirements
CDSS allows for better cognitive analysis of multiple factors for clinicians in a way that they may not be able to perform manually.
Intervention Design
What sets this study apart is its comprehensive approach to implementation and evaluation. The intervention combined three key elements:
1. A sophisticated CDSS tool capable of analyzing over 160 potential diabetes treatments
2. A team-based care model incorporating pharmacists
3. Direct patient engagement in treatment decisions
Study Results and Clinical Impact
The results demonstrated meaningful clinical improvements. Patients whose providers followed the CDSS recommendations either fully or partially showed statistically significant reductions in HbA1c levels:
Initial average HbA1c: 10.1
Final average HbA1c: 8.0
Patients receiving treatment based on full CDSS recommendations showed a 17.8% reduction in HbA1c
Those receiving partially CDSS-aligned treatments showed a 20.96% reduction in HbA1c
These improvements were more substantial than those observed in the control group, suggesting that the combination of CDSS and team-based care can effectively address clinical inertia.
Provider Adoption and Feedback
Importantly, the study also examined the human factors crucial for successful CDSS implementation:
Over 80% of physicians believed the CDSS would improve patient outcomes
65% reported expanded knowledge of medication options
85% expressed comfort with CDSS-recommended treatments
More than half indicated the experience would lead to permanent changes in their practice
The pharmacists involved in the study reported that the CDSS improved:
Formulary transparency
Treatment option awareness
Patient engagement
Communication between care team members
Implementation Challenges and Lessons
The study identified several key challenges that must be addressed for successful CDSS deployment:
1. Integration with existing electronic health record systems (this is a huge problem and with Epic’s and Cerner’s (now Oracle Health) monopolies on the clinical setting, they have great control over which systems to integrate.
2. Workflow optimization to minimize time burden (workflows must be built into the EHR environment)
3. Regular updates to medication databases and formularies (CDSS systems must be readily updated as new evidence emerges)
4. Adequate training for all team members (turn-over and training are two major barriers to success)
5. Management of technical issues and system responsiveness (technical issue support is always present in software-based systems)
Conclusion
As health care continues to embrace technological solutions, this study serves as a model for how CDSS should be validated and implemented. The results demonstrate that achieving real clinical improvements requires more than just sophisticated algorithms – it demands careful attention to workflow integration, team dynamics, and patient engagement. Future CDSS developments would benefit from following similar comprehensive validation approaches to ensure they deliver on their promise of improving patient care. Both clinical effectiveness and clinician-usability are critical factors to consider in CDSS implementations.
This type of thorough clinical validation, while resource-intensive, provides the evidence needed to support broader adoption of CDSS in clinical practice and helps ensure that technological innovations translate into meaningful improvements in patient outcomes.
Let me know what you think!
The proprietary EHR (EPIC/Cerner etc) are a major problem for improvements in health care as well as the availability of care. The time and finanacial resources required to implement and maintain these systems is huge. Health care workers spend an inordinate amount of time interacting with their computers rather than with their patients. But the issue for today is that health care workers are no longer able to fluidly move from one institution or employer to another. A common interface or an inter face that was more intuitive using Natural Language.
We are never going to go back to a paper chart or paper documentation but the current systems do not enhance the bedside experience and therefore they do not enhance the patient experience.