Browsing by Subject "Decision Support Systems, Clinical"
Now showing 1 - 5 of 5
- Results Per Page
- Sort Options
Item Automating Clinical Decision Support to Improve Pediatric Hypertension Identification: Evaluation of Diagnostic Validity and Computational Time of Two Algorithms(2017-01-17) Doney, Analise; Bowen, Michael; Adamson, Brian; Sanders, Joanne; Vertilus, Shawyntee; Menzies, Christopher; Gheen, Taylor; Bhat, Deepa; Fish, Jason; Skinner, Celette Sugg; Turer, Christy BolingHypertension (HTN) in children is underdiagnosed. This is concerning. Rates of HTN are increasing with the rise in childhood obesity. In children, diagnosis of HTN requires three discrete elevations in historical blood-pressure (BP) percentiles (%). Calculating BP% requires valid heights and ages that change from visit to visit, and BP%s are neither calculated nor stored in electronic health records (EHRs). Current methods used by providers to circumvent these challenges may contribute to under diagnosis of pediatric HTN. EHR-enabled algorithms may improve rates of HTN diagnosis by automating discovery of historical BP% elevations. The study aim was to determine whether an EHR-enabled algorithm to diagnose HTN is more diagnostically valid than methods currently used by providers. Data were from a retrospective cohort of 52,828 3-18 year-old children followed up to 6 years in primary care. The 1st algorithm, termed current-observation-carried backwards (COCB), identified the BP% threshold for HTN using current-visit height/age, and applied that threshold to historical visits to determine HTN (defined as 3 elevations in BP% ≥95). COCB imitated mental processes currently used by providers, because retrospectively calculating visit-specific thresholds during patient visits is not feasible. The 2nd algorithm, termed “smart data elements” (SDEs), exported historical-visit ages, heights, and BPs from the EHR into a database to calculate historical BP% thresholds for HTN—thus providing visit-specific BP%s. The 2nd algorithm is the ideal way to determine HTN. The study hypothesis was that the second/SDE algorithm would be more diagnostically valid than the first/COCB. Diagnostic validity for determining HTN was determined by calculating sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NPV). Descriptive statistics summarized diagnostic validity of the algorithms. Variation in diagnostic validity by height and body-mass-index percentile (BMI%) were explored. The sensitivity of COCB (vs. SDE) for detecting HTN was 82.2%. COCB’s specificity surpassed 99.8%. PPV and NPV were >95%. Diagnostic validity differed by height and BMI%: as height and BMI% increased, sensitivity increased and PPV decreased. Study data indicate that the current method used by many pediatricians to diagnose HTN (COCB) misses up to 1 in 5 children with HTN. Automating identification of three discrete BP% elevations using the SDE algorithm may improve rates of pediatric HTN diagnosis.Item Bringing machine learning to the bedside: focusing clinical decision support with predictive modeling(2019-01-18) Basit, MujeebItem Decision Fatigue in Primary Care Opioid Prescribing(2020-05-01T05:00:00.000Z) Hughes, Jordan Gregory; Kandil, Enas; Reed, W. Gary; Greilich, PhilipBACKGROUND: Decision fatigue -- a psychological phenomenon describing the depletion of mental resources as one makes a series of decisions -- affects primary care physicians as they treat patients and prescribe medications throughout the course of the clinical day. This results in more inappropriate treatments being ordered as the day goes on. Because the United States faces an unprecedented epidemic of opioid abuse, we must understand the extent to which decision fatigue affects opioid prescribing, as the prescription of these drugs has been associated with both long-term use and overdose deaths. Additionally, we are unaware of the effect various national interventions to stem the tide of the opioid epidemic have had on decision fatigue in opioid prescribing. LOCAL PROBLEM: We are unaware of whether decision fatigue is playing a role in opioid prescribing, and if it is, how great the variation in prescription likelihood is throughout the clinical day in UT Southwestern's primary care clinics. The aim of this study is to measure PCPs' varying likelihoods of prescribing opioids throughout the clinical day, before and after major interventions were implemented to combat the epidemic, as this can serve as an indication of both the presence of decision fatigue and the impact of concerted interventions. METHODS: We used the years 2014 and 2017 to represent the pre- and post-intervention periods for study, as many major interventions to combat the opioid epidemic took place in 2016. Next, we analyzed the percentage of appointments in which opioids were prescribed in each hour of physicians' clinical days, at three exclusively primary care clinics at UT Southwestern. Scheduled appointment times were used as substitutes for visit times. We then excluded patients with cancer and those who had surgery within six weeks of an appointment, in order to minimize the number of appointments in which opioids may be prescribed by clear clinical indication. Finally, we employed logistic regression analysis to determine the predictive relationship between appointment time and opioid prescriptions, using physicians' prescription rates in their first clinic hour as the reference for calculating hourly odds ratios in each year. INTERVENTIONS: New legislation, updated healthcare guidelines, and national media coverage in 2016, including: the CDC's "Guideline for Prescribing Opioids for Chronic Pain—United States", the FDA's "General Principles for Evaluating the Abuse Deterrence of Generic Solid Oral Opioid Drug Products: Guidance for Industry, 2016", the Comprehensive Addiction and Recovery Act of 2016, the Surgeon General's call to providers to end the opioid crisis, and the opioid crisis becoming top health news story of 2016. RESULTS: 34,972 clinic visits in 2014 and 42,313 clinic visits in 2017 met our inclusion criteria. In 2014, patients were prescribed an opioid at 5.34% of all primary care appointments, while in 2017, they were prescribed at a rate of 4.34% (2014 vs 2017 OR=1.290; 95% CI, 1.217-1.367). While the overall rates of opioids decreased from 2014 to 2017, the hourly likelihoods of patients being prescribed opioids steadily increased throughout the clinical day in both years (p<0.01 in 2014 and 2017). In fact, each year had hours in which physicians' opioid prescription odds ratio was over 1.6, when compared to their first clinical hour. DISCUSSION: While there was a significant decrease in the overall likelihood of being prescribed opioids in 2017 compared to 2014, the variation in hourly prescription likelihoods is similar in both years. The results show that while interventions to combat the opioid epidemic were successful in reducing the overall amount of opioids prescribed, they had minimal impact on the effect of decision fatigue. Clinical decision support tools, integrated into the electronic medical record, have been proven to reduce the clinical variation that indicates the presence of decision fatigue. In light of this, this project should be continued, and decision fatigue measured, after implementation of such tools.Item Improving the Error Review Process for Incident Reports at UT Southwestern Through the Use of a Standardized Taxonomy Tool(2021-04-22) Chan, Christopher; Reed, W. Gary; Lynch, Isaac; Greilich, PhilipBACKGROUND: At UT Southwestern and in other hospital systems, medical errors are often the result of latent system errors. These errors may remain unreported due to a variety of reasons. When they are reported, the defect analysis protocols triggered by error review are often inadequate for identification and correction of the underlying system defect. The aim of this project was to improve the process for classifying medical errors at UT Southwestern by testing the use of a standardized taxonomy tool, refining the tool to reduce inter-rater variability, and incorporating its use into the existing workflow for real-time error review. Of particular interest was the subset of medical errors related to OR-ICU handover. METHODS/INTERVENTION: The taxonomy tool used in this project was created by Dr. Isaac Lynch, cardiac anesthesia and intensive care faculty at UT Southwestern. It is a form that allows categorization of several important parameters described by both the World Health Organization's 2005 "Draft Guidelines for Adverse Event Reporting and Learning Systems" and 2014 "Minimal Information Model for Patient Safety Incident Reporting" guidelines. For example, the tool allows for designation of an error as being "administrative," "medication," or "handover-related." Properly trained Clinical Safety personnel can use the taxonomy tool to quickly review incident reports in real-time as they are submitted. For this project, all ICU incident reports submitted through the UT Southwestern Clements University hospital reporting system in 2015, a total of 1317 reports, were reviewed and classified according to the taxonomy tool. The collected data was analyzed using REDCapTM built-in statistical modeling tools for hazard identification and trends. Based on this initial trial run, the taxonomy tool was refined to improve functionality and reduce inter-rater variability that may occur when different people use the tool. Analysis of the data also showed how taxonomic categorization can highlight system errors that are in need of targeted intervention. Finally, a process map was created to illustrate how use of this taxonomy tool can be incorporated into the existing error review workflow at UT Southwestern. RESULTS: Analysis of the incident reports from 2015 and classification according to the taxonomy tool revealed some specific areas in need of process improvement. For example, there were a significant number of reports related to inappropriate specimen labeling and also many reports describing patient pressure ulcers. Use of the taxonomy tool enabled identification and classification of these medical errors. Future, in-depth analysis can then be used to inform targeted intervention. Notable statistics from the analysis are that 46% of the reports described errors occurring during perioperative care, 18% were medication errors, and 18% were diagnostic errors. To further classify the factors contributing to these different errors, nearly 52% of the total incident reports were latent system errors, 58% could be attributed to staff error, and 45% had some component of patient cause. Of note, a significant 70% of the latent system errors were equipment-related. Also, only 5.5% of the reports described errors that were probably related to handover, with 78.4% unlikely to be related to handover. Finally, 61% of the medical errors caused temporary harm to the patient, with 1.9% describing errors that contributed to patient death. CONCLUSION: This project demonstrated that it is both practical and helpful to use a standardized taxonomy tool for routine, real-time error review. Review of a single incident report and completion of the corresponding taxonomy form takes an average of 7 minutes for trained personnel so there will not be a significant additional work burden for Clinical Safety specialists. The data collected in this project has been helpful for identifying specific system errors, and classification of the types of error has utility in determining the best intervention strategies. The gradual accumulation of data can also be used for trend identification and epidemiological studies. In terms of next steps, the taxonomy tool, which has already been modified based on this initial test run, will undergo further trial testing to improve inter-rater reliability. This will involve training other medical students in the tool's proper use and having them repeat error review on a subset of the 2015 ICU reports, before then continuing the analysis with ICU reports from subsequent years. Future projects will also focus on integrating use of the taxonomy tool into the existing error review workflow, developing a dashboard module for real-time trend analysis, and enabling human languages algorithm functionality. The vision is for the taxonomy tool to be employed by Clinical Safety personnel in real-time to generate useable classification data for fixing latent system errors.