Decision support systems are increasingly embedded in healthcare operations, from electronic health records and medication alerts to predictive-AI tools. The evidence shows measurable gains in prescribing, testing, documentation, vaccination, and selected risk-management tasks, but effects vary by workflow, alert design, and implementation.
Contents
- Adoption and reach
- Effects on care processes
- Medication safety and prescribing
- Risk stratification and chronic disease
- Prevention, diagnosis, and public health
- Workflow, alerts, and implementation
Adoption and reach
Decision support is not a niche feature of modern health information systems. By 2017, more than 90% of U.S. hospitals and 80% of U.S. clinics had electronic health records with some form of clinical decision support. This older estimate describes broad availability, not uniform use or effectiveness. AHRQ, “Clinical Decision Support Systems”
Predictive-AI adoption among U.S. hospitals rose from 66% in 2023 to 71% in 2024. In 2024, 80% used predictive AI sourced from their EHR developer, 52% used third-party-developed tools, and 50% used self-developed tools. These categories can overlap because a hospital may use more than one type of system. HealthIT, “Hospital Trends in the Use, Evaluation, and Governance of Predictive AI, 2023-2024”
Use is expanding beyond direct clinical recommendations. U.S. hospital use of predictive AI for billing automation increased from 36% in 2023 to 61% in 2024, while use for scheduling rose from 51% to 67%. Among hospitals using third-party or self-developed predictive AI, use to identify high-risk outpatients for follow-up increased by 51 percentage points between 2023 and 2024. HealthIT, “Hospital Trends in the Use, Evaluation, and Governance of Predictive AI, 2023-2024”
Governance is also distributed. In 2024, three-quarters of U.S. hospitals said more than one party was responsible for evaluating predictive-AI models, and more than half said three or more parties were responsible. Adoption statistics therefore need to be read alongside accountability and evaluation practices. HealthIT, “Hospital Trends in the Use, Evaluation, and Governance of Predictive AI, 2023-2024”
Effects on care processes
The strongest consistent evidence concerns care processes rather than mortality. A review of literature published from 1976 through 2010 found improved preventive-service ordering or completion with an odds ratio of 1.42 (95% CI 1.27–1.58) across 25 studies. Clinical-study ordering improved with an odds ratio of 1.72 (95% CI 1.47–2.00) across 20 studies, while therapy prescribing improved with an odds ratio of 1.57 (95% CI 1.35–1.82) across 46 studies. AHRQ, “Enabling Health Care Decisionmaking Through Clinical Decision Support and Knowledge Management”
That review also shows an important evidence limitation: only 29 of 148 randomized clinical decision-support trials, or 19.6%, assessed clinical outcomes. Only 22 trials, or 14.9%, assessed costs. Many studies therefore measured whether recommended actions changed, not whether patients lived longer, recovered faster, or incurred lower total costs. AHRQ, “Enabling Health Care Decisionmaking Through Clinical Decision Support and Knowledge Management”
A later meta-analysis of 122 controlled trials included 1,203,053 patients and 10,790 providers. Desired-care delivery increased by an absolute 5.8% (95% CI 4.0%–7.6%). By process, prescribing improved 4.4% (95% CI 2.6%–6.2%) across 68 trials, test ordering improved 6.8% (95% CI 4.9%–8.6%) across 30 trials, documentation adherence improved 7.1% (95% CI 5.4%–8.9%) across 25 trials, and vaccination improved 5.9% (95% CI 4.1%–7.7%) across 12 trials. These estimates cover trials published through August 2019. “Computerised clinical decision support systems and absolute improvements in care”
| Care process | Absolute improvement | Evidence base |
|---|---|---|
| Prescribing | 4.4% | 68 trials |
| Test ordering | 6.8% | 30 trials |
| Documentation adherence | 7.1% | 25 trials |
| Vaccination | 5.9% | 12 trials |
Medication safety and prescribing
Medication-related decision support can change clinician actions without necessarily changing short-term patient outcomes. In a 2010–2012 hospital randomized trial, the system generated 28,394 medication-related reminders, with a median of three reminders per patient per hospital stay. Reminder resolution was 38.0% in the decision-support group versus 33.7% in controls, an odds ratio of 1.21 (95% CI 1.11–1.32). In-hospital all-cause mortality did not differ significantly: odds ratio 0.95 (95% CI 0.77–1.17). “Effectiveness of a Hospital-Based Computerized Decision Support System on Clinician Recommendations and Patient Outcomes”
In an acute-kidney-injury alert trial conducted from 2016 to 2022, potentially nephrotoxic medication was discontinued within 24 hours for 61.1% of alert patients versus 55.9% receiving usual care. The 14-day composite of kidney-injury progression, dialysis, or death occurred in 23.1% of alert patients and 25.3% of usual-care patients (RR 0.92, 95% CI 0.83–1.01). “A randomized clinical trial assessing the effect of automated medication-targeted alerts on acute kidney injury outcomes”
Antimicrobial support also produced process and financial changes. Between May and August 2004, an antimicrobial-management team intervened on 16% of 2,237 intervention-arm inpatients compared with 8% of 2,270 controls. The computerized system reduced hospital antimicrobial expenditure by 23%, or $84,194, equivalent to $37.64 per patient. “Impact of a Computerized Clinical Decision Support System on Reducing Inappropriate Antimicrobial Use”
Risk stratification and chronic disease
In a community-health-center trial from September 2018 through March 2020, 70 clinics were randomized: 42 to clinical decision support and 28 to control. The cardiovascular-risk system was used in 19.8% of 91,988 eligible intervention-clinic encounters. Mean baseline total cardiovascular risk was 16.6% in control-clinic patients and 15.6% in intervention-clinic patients. Among patients at high baseline risk, mean reversible cardiovascular risk decreased by 4.4 percentage points when the system was used versus 2.7 points in controls. “Effect of Clinical Decision Support at Community Health Centers on the Risk of Cardiovascular Disease”
The HIV-clinic evidence illustrates how alert design can affect follow-up. From September 2007 through September 2008, mean monthly CD4-count increase was 5.3 cells/mm³ with interactive alerts versus 3.2 cells/mm³ with static alerts. Six-month suboptimal follow-up occurred at 20.6 versus 30.1 events per 100 patient-years, respectively. After a suboptimal-follow-up alert, median time to the next scheduled appointment was 1.71 months with interactive alerts versus 3.48 months with static alerts. After a toxicity alert, the median was 2.79 months with interactive alerts versus more than 6 months with static alerts. “Successful Outcomes of a Clinical Decision Support System in an HIV Practice”
Prevention, diagnosis, and public health
Results in preventive care are mixed and highly outcome-specific. In a predominantly rural cancer-screening trial conducted from August 2018 through March 2020, 9,543 of 25,198 overdue patients, or 37.9%, were up to date on the composite screening endpoint after 12 months. Adjusted 12-month composite screening status was 38.1% with decision support, 36.5% with usual care, and 34.4% with decision support plus shared decision-making tools. The adjusted odds ratio for decision support versus usual care was 1.07 (97.5% CI 0.89–1.29; p=0.38). “Clinical decision support with or without shared decision making to improve preventive cancer care”
For lung-cancer screening in the same trial, adjusted up-to-date status was 30.6% with decision support, 22.1% with usual care, and 22.4% with decision support plus shared decision-making. The adjusted odds ratio for decision support versus usual care was 1.55 (97.5% CI 0.84–2.86; p=0.099). The point estimate favored decision support, but the reported interval and p-value do not establish a statistically significant difference under conventional thresholds. “Clinical decision support with or without shared decision making to improve preventive cancer care”
Documentation can respond strongly to targeted prompts. In a November 2009–November 2010 problem-list trial, providers accepted 41.1% of 17,043 clinical alerts. The alerts increased provider documentation of study problems with an adjusted odds ratio of 3.4 (p<0.001). “Improving completeness of electronic problem lists through clinical decision support”
In a Kenyan HIV trial conducted from September 2012 through January 2014, appropriate action on immunological treatment failure occurred for 54% of intervention patients versus 30% of controls. The intervention raised the likelihood of appropriate treatment-failure action to an adjusted odds ratio of 3.18 (95% CI 1.02–9.87). “Effect of a clinical decision support system on early action on immunological treatment failure in patients with HIV in Kenya”
Workflow, alerts, and implementation
Implementation determines how much of a system’s potential reaches practice. A taxonomy of 74 decision-support systems from 58 randomized trials found that 38% were patient-directed preventive or health-behavior systems, while 18% targeted inpatients with online clinician decision support and direct online execution of recommendations. Extra staffing for decision-support input or output was required by 84% of the 74 systems. The taxonomy covered randomized trials published from 1998 through 2003. AHRQ, “A taxonomic description of computer-based clinical decision support systems”
A systematic review of 100 controlled trials found improved practitioner performance in 62 of 97 studies that measured it. Automatically prompted decision support improved performance in 73% of trials, compared with 47% when users had to activate the system. Studies in which the authors also developed the software reported success in 74% of trials, versus 28% when they did not develop it. These results come from literature available through 2004 and should not be treated as current adoption estimates. AHRQ, “Clinical Decision Support”
In the iCPR randomized trial, adoption was 57.5% among intervention users and acceptance was 42.4%. The intervention reduced antibiotic ordering for streptococcal pharyngitis by 9.2% (p=0.008). In a home-care medication-complexity trial, hospitalization occurred in approximately 20% of intervention patients versus 21% of controls, while emergency-department use was 16.5% versus 16.7%. “A Framework for Usable and Effective Clinical Decision Support”; “Outcomes of clinical decision support and correlates of CDS use for home care patients”
Alert burden and relevance remain practical constraints. From January 2009 through December 2011, appropriate override rates for ICU medication alerts ranged from 94% for allergy alerts to 27% for renal alerts. Drug-drug-interaction alert overrides were appropriate in 84% of cases, and geriatric-alert overrides were appropriate in 57%. “Evaluation of medication-related clinical decision support alert overrides in the intensive care unit”
Outcome effects also differ by alert and condition. A real-time pneumonia decision-support trial conducted from 2017 through 2021 reported 38% lower severity-adjusted 30-day all-cause mortality after deployment. By contrast, in a randomized severe-sepsis alert evaluation, new antibiotic orders within three hours occurred in 35% of alert patients versus 37% of controls (p=0.53). In a randomized deterioration-alert trial, ICU transfer occurred in 17.8% of intervention patients versus 18.2% of controls, and hospital mortality was 7.3% with alerts versus 7.7% without alerts. “A Pragmatic, Stepped-Wedge, Cluster-controlled Clinical Trial of Real-Time Pneumonia Clinical Decision Support”; “Electronic health record-based clinical decision support alert for severe sepsis”; “A RANDOMIZED TRIAL OF REAL-TIME AUTOMATED CLINICAL DETERIORATION ALERTS”