Clinical Decision Support Tool Research Articles

Using a Built-in Clinical Decision Support to Improve Phosphate Repletion Practice: A Quasi-Experimental Study

BackgroundInpatient serum phosphate replacement is common, but there is great variability in replacement practice, which leads to overuse. Electronic medical record (EMR) interventions with clinical decision support (CDS) can be effective tools to guide clinicians towards best clinical practices. Our objective was to utilize CDS tools to reduce overuse of hypophosphatemia corrections at a large safety-net healthcare system. MethodsOur first intervention involved enhancing an existing order set for phosphate repletion by incorporating CDS to guide appropriate repletion orders based on deficit severity and simplifying ordering. The second intervention was a best practice advisory (BPA) that triggered when an intravenous (IV) phosphate repletion was ordered for a patient with mild to moderate phosphate deficiency without an existing nil per os (NPO) order. The primary outcome measure was the number of patients with mild and moderate hypophosphatemia receiving IV replacement without NPO orders per 1000 patient days. ResultsAcross all hospitals, rate of IV replacement in patients with mild to moderate hypophosphatemia (1.0 to 1.9 mg/dL) without NPO orders decreased from 7.22 to 3.40 per 1000 patient days (53.0% reduction, p<0.001), while the oral replacements in this population increased from 6.39 to 8.87 (38.8% increase, p<0.001). For patients with phosphate levels ≥2.0, IV replacements decreased from 10.66 to 5.36 (49.8% reduction, p<0.001) and oral replacements decreased 19.78 to 16.69 (15.6% reduction, p<0.01). ConclusionsOur intervention successfully reduced inpatient IV phosphate replacements by 53% in patients with mild to moderate hypophosphatemia using a two-pronged electronic health record intervention across a large safety net setting.

RICD: Russian Intensive Care Dataset

In the era of healthcare digital transformation, the scientific community faces the need for structured and available datasets for research and technological projects in the field of artificial intelligence, related to the development of new diagnostic and treatment methods.Objective: to develop a dataset containing anonymized medical data of all patients treated at the Federal Research and Clinical Center of Intensive Care Medicine and Rehabilitology (FRCCR), and provide access for doctors and scientists of FRCCR and other centers to structured patient data for subsequent analysis and research. Materials and Methods. The FRCCR medical information system and the tools «Asclepius», PL/SQL, Microsoft Office Excel, Power Query M, Microsoft PowerBI, Open data editor, and Python were used for data collection and representation. To provide open access to the dataset and protect the personal data of patients, the information was anonymized.Results. We introduce the RICD (Russian Intensive Care Dataset, https://fnkcrr-database.ru/) — the first dataset of intensive care patients in the Russian Federation, developed at FRCCR based on advanced principles and methods used in international open database projects — «eICU Program» from Philips Healthcare, «MIMIC-IV», and «MIMIC-III». The developed dataset contains information on 7,730 hospitalizations of 5,115 patients (including readmissions), covering data from 3,291 hospitalizations in the intensive care units (ICUs). The total number of records in the RICD exceeds 14 million. The RICD presents medical-anthropometric data, patient movement within the institution, diagnoses, information on therapy provided, results of laboratory tests, scale assessments, and outcomes of hospitalization. RICD also contains data on several vital parameters collected from bedside monitors and other equipment of ICUs, with up to 10 evaluations per hour.Conclusion. The RICD allows for in-depth analysis and research of clinical practices in intensive care, enabling the development of clinical decision support tools and the application of machine learning methods to enhance diagnostic tools and improve patient outcomes. With its accessibility and detailed data structure, the dataset serves as a valuable tool for both scientific research and practical applications in intensive care.

Towards objective and systematic evaluation of bias in artificial intelligence for medical imaging.

Artificial intelligence (AI) models trained using medical images for clinical tasks often exhibit bias in the form of subgroup performance disparities. However, since not all sources of bias in real-world medical imaging data are easily identifiable, it is challenging to comprehensively assess their impacts. In this article, we introduce an analysis framework for systematically and objectively investigating the impact of biases in medical images on AI models. Our framework utilizes synthetic neuroimages with known disease effects and sources of bias. We evaluated the impact of bias effects and the efficacy of 3 bias mitigation strategies in counterfactual data scenarios on a convolutional neural network (CNN) classifier. The analysis revealed that training a CNN model on the datasets containing bias effects resulted in expected subgroup performance disparities. Moreover, reweighing was the most successful bias mitigation strategy for this setup. Finally, we demonstrated that explainable AI methods can aid in investigating the manifestation of bias in the model using this framework. The value of this framework is showcased in our findings on the impact of bias scenarios and efficacy of bias mitigation in a deep learning model pipeline. This systematic analysis can be easily expanded to conduct further controlled in silico trials in other investigations of bias in medical imaging AI. Our novel methodology for objectively studying bias in medical imaging AI can help support the development of clinical decision-support tools that are robust and responsible.

Use of the iPRISM webtool in a learning community to assess implementation context and fit of a novel clinical decision support tool for physical therapy triage in acute care hospitals.

The iPRISM webtool is an interactive tool designed to aid the process of applying the Practical, Robust Implementation and Sustainability Model (PRISM) for the assessment of and fit with context. A learning community (LC) is a multidisciplinary group of partners addressing a complex problem. Our LC coproduced the Physical TheraPy frEqueNcy Clinical decIsion support tooL (PT-PENCIL) to guide the use of physical therapist services in acute care hospitals. To describe our LC's activities to co-produce the PT-PENCIL, use of the iPRISM webtool to assess its preimplementation context and fit, and develop a multicomponent implementation strategy for the PT-PENCIL. A descriptive research design. Three tertiary care hospitals. Thirteen LC partners: six clinical physical therapists, three rehabilitation managers, three researchers, and a bioinformaticist. Not applicable. Using the iPRISM webtool, expected fit of the PT-PENCIL was rated 1 (not aligned) to 6 (well aligned) for each PRISM domain and expected reach, effectiveness, adoption, implementation, and maintenance were rated 1 (not likely at all) to 6 (very likely). Discrete implementation strategies were identified from the Expert Recommendations for Implementing Change. The process spanned 18 meetings over 8 months. Ten LC partners completed the iPRISM webtool. PRISM domains with the lowest expected alignment were the "implementation and sustainability infrastructure" (mean = 4.7 out of 6; range = 3-6) and the "external environment" (mean = 4.9 of 6; range = 4-6). Adoption was the outcome with the lowest expected likelihood (mean = 4.5 out of 6; range = 1-6). Six discrete implementation strategies were identified and combined into a multicomponent strategy. Within a LC, we used existing implementation science resources to co-produce a novel clinical decision support tool for acute care physical therapists and develop a strategy for its implementation. Our methodology can be replicated for similar projects given the public availability of each resource used.

Computed tomography-based structural rigidity analysis can assess tumor- and treatment-induced changes in rat bones with metastatic lesions

BackgroundBreast cancer (BrCa) is a predominant malignancy, with metastasis occurring in one in eight patients, nearly half of which target the bone, leading to serious complications such as pain, fractures, and compromised mobility. Structural rigidity, crucial for bone strength, becomes compromised with osteolytic lesions, highlighting the vulnerability and increased fracture risk in affected areas. Historically, two-dimensional radiographs have been employed to predict these fracture risks; however, their limitations in capturing the three-dimensional structural and material changes in bone have raised concerns. Recent advances in CT-based Structural Rigidity Analysis (CTRA), offer a promising, more accurate non-invasive 3D approach. This study aims to assess the efficacy of CTRA in monitoring osteolytic lesions’ progression and response to therapy, suggesting its potential superiority over existing methodologies in guiding treatment strategies.MethodsTwenty-seven female nude rats underwent femoral intra-medullary inoculation with MDA-MB-231 human breast cancer cells or saline control. They were divided into Control, Cancer Control, Ibandronate, and Paclitaxel groups. Osteolytic progression was monitored weekly using biplanar radiography, quantitative computed tomography (QCT), and dual-energy X-ray absorptiometry (DEXA). CTRA was employed to predict fracture risk, normalized using the contralateral femur. Statistical analyses, including Kruskal-Wallis and ANOVA, assessed differences in outcomes among groups and over time.ResultsBiplanar radiographs showed treatment benefits over time; however, only certain time-specific differences between the Control and other treatment groups were discernible. Notably, observer subjectivity in X-ray scoring became evident, with significant inter-operator variations. DEXA measurements for metaphyseal Bone Mineral Content (BMC) did not exhibit notable differences between groups. Although diaphyseal BMC highlighted some variance, it did not reveal significant differences between treatments at specific time points, suggesting a limited ability for DEXA to differentiate between treatment effects. In contrast, the CTRA consistently demonstrated variations across different treatments, effectively capturing bone rigidity changes over time, and the axial- (EA), bending- (EI), and torsional rigidity (GJ) outcomes from the CTRA method successfully distinguished differences among treatments at specific time points.ConclusionTraditional approaches, such as biplanar radiographs and DEXA, have exhibited inherent limitations, notably observer bias and time-specific inefficacies. Our study accentuates the capability of CTRA in capturing real-time, progressive changes in bone structure, with the potential to predict fractures more accurately and provide a more objective analysis. Ultimately, this innovative approach may bridge the existing gap in clinical guidelines, ushering in enhanced Clinical Decision Support Tool (CDST) for both surgical and non-surgical treatments.

Cross‐sectional evaluation of a clinical decision support tool to identify medication‐related problems at discharge from the acute care setting

AbstractBackgroundThere are many reported pharmacist‐led transitions of care (TOC) programs to address medication‐related problems (MRP) at discharge from the acute care setting. Most have identified time and labor resources as significant limitations. This study aims to assess the effectiveness of a medication risk score (MRS)‐driven clinical decision support system (CDSS) in identifying actionable MRPs and improving medication safety in the acute care discharge TOC setting.MethodsA cross‐sectional analysis was conducted in a cohort of 481 subjects discharged from the acute care setting. The MRS‐CDSS was utilized to identify MRPs and provide recommendations for risk reduction. The distribution of MRPs, recommendations, and their associations with MRS severity were analyzed. Additionally, the potential reduction in MRS per subject and its correlation with MRS severity were examined.ResultsThe median MRS reduction per subject was 2 points, while high/severe‐risk patients showed a median potential reduction of 7 points. Among the identified MRPs (n = 691), drug interaction, drug use without indication, and adverse drug reaction accounted for 89.7% of all MRPs. The top three recommendations, discontinue medication, change the time of administration, and start alternative therapy, represented 94.1% of all recommendations. Stratified analysis by MRS category revealed a significant increase in adverse drug reaction MRPs and recommendations to discontinue medications with higher MRS severity. The results were consistent with previous outpatient studies, supporting the MRS‐CDSS's ability to enhance medication safety.ConclusionThis study demonstrates that the MRS‐CDSS effectively identifies actionable MRPs and has the potential to substantially reduce overall pharmacotherapy regimen risk when applied during acute care discharge TOC. The findings support implementable recommendations directed at patient safety and the allocation of health care resources to high‐risk patients for maximum benefit.

Enhancing patient safety: leveraging artificial intelligence-powered electronic medical records for effective drug-drug interaction nudge in real-world prescribing practices

Background: Concurrent prescriptions of various medications may lead to unfavorable and unanticipated potential drug-drug interactions. Hence, the elimination of drug-drug interactions is a key aspect of delivering a coherent treatment regime. In response to this concern, HealthPlix, one of India's largest AI-powered electronic medical record providers, introduced a drug-drug interaction nudge feature in June 2022, providing a proactive solution for physicians to address potential interactions between incompatible drugs. This study aimed to elucidate the role of electronic medical records in identifying and managing drug interactions and the advantages of interaction nudges for doctors in prescribing appropriate medications. Methods: An observational retrospective study was conducted using data obtained from HealthPlix, containing two or more drugs, written for patients older than 18 years. Results: In an average of 1.9 million patient visits analyzed, the interaction visits were observed to be 1.2 million. An average of 185,745 interactions were observed during the study period. For all observed interactions, an average of 72,383 molecules were removed. These results provide insights into the efficiency of HealthPlix in abrogating interactions and illustrate the tangible benefits of nudges in modifying prescription practices. Conclusions: The above results illustrate the effectiveness of drug-drug interaction nudges as a clinical decision support tool integrated into HealthPlix, marking a significant advancement in Indian healthcare. This unique feature contributes to reducing the frequency of potent drug interactions, showcasing its potential to enhance patient safety and improve the quality of healthcare delivery.

Optimising a clinical decision support tool to improve chronic kidney disease management in general practice

BackgroundEarly identification and treatment of chronic disease is associated with better clinical outcomes, lower costs, and reduced hospitalisation. Primary care is ideally placed to identify patients at risk of, or in the early stages of, chronic disease and to implement prevention and early intervention measures. This paper evaluates the implementation of a technological intervention called Future Health Today that integrates with general practice EMRs to (1) identify patients at-risk of, or with undiagnosed or untreated, chronic kidney disease (CKD), and (2) provide guideline concordant recommendations for patient care. The evaluation aimed to identify the barriers and facilitators to successful implementation.MethodsFuture Health Today was implemented in 12 general practices in Victoria, Australia. Fifty-two interviews with 30 practice staff were undertaken between July 2020 and April 2021. Practice characteristics were collected directly from practices via survey. Data were analysed using inductive and deductive qualitative analysis strategies, using Clinical Performance - Feedback Intervention Theory (CP-FIT) for theoretical guidance.ResultsFuture Health Today was acceptable, user friendly and useful to general practice staff, and supported clinical performance improvement in the identification and management of chronic kidney disease. CP-FIT variables supporting use of FHT included the simplicity of design and delivery of actionable feedback via FHT, good fit within existing workflow, strong engagement with practices and positive attitudes toward FHT. Context variables provided the main barriers to use and were largely situated in the external context of practices (including pressures arising from the COVID-19 pandemic) and technical glitches impacting installation and early use. Participants primarily utilised the point of care prompt rather than the patient management dashboard due to its continued presence, and immediacy and relevance of the recommendations on the prompt, suggesting mechanisms of compatibility, complexity, actionability and credibility influenced use. Most practices continued using FHT after the evaluation phase was complete.ConclusionsThis study demonstrates that FHT is a useful and acceptable software platform that provides direct support to general practice in identifying and managing patients with CKD. Further research is underway to explore the effectiveness of FHT, and to expand the conditions on the platform.

Online interpretable dynamic prediction models for clinically significant posthepatectomy liver failure based on machine learning algorithms: A retrospective cohort study.

Posthepatectomy liver failure (PHLF) is the leading cause of mortality in patients undergoing hepatectomy. However, practical models for accurately predicting the risk of PHLF are lacking. This study aimed to develop precise prediction models for clinically significant PHLF. A total of 226 patients undergoing hepatectomy at a single center were recruited. The study outcome was clinically significant PHLF. Five pre- and postoperative machine learning (ML) models were developed and compared with four clinical scores, namely, the MELD, FIB-4, ALBI, and APRI scores. The robustness of the developed ML models was internally validated using 5-fold cross-validation by calculating the average of the evaluation metrics and was externally validated on an independent temporal dataset, including the area under the curve (AUC) and the area under the precision‒recall curve (AUPRC). SHapley Additive exPlanations analysis was performed to interpret the best performance model. Clinically significant PHLF was observed in 23 of 226 patients (10.2%). The variables in the preoperative model included creatinine, total bilirubin, and Child‒Pugh grade. In addition to the above factors, the extent of resection was also a key variable for the postoperative model. The pre- and postoperative artificial neural network (ANN) models exhibited excellent performance, with mean AUCs of 0.766 and 0.851, respectively, and mean AUPRC values of 0.441 and 0.645, whereas the MELD, FIB-4, ALBI, and APRI scores reached AUCs of 0.714, 0.498, 0.536 and 0.551, respectively, and AUPRC values of 0.204, 0.111, 0.128 and 0.163, respectively. In addition, the AUCs of the pre- and postoperative ANN models were 0.720 and 0.731, respectively, and the AUPRC values were 0.380 and 0.408, respectively, on the temporal dataset. Our online interpretable dynamic ML models outperformed common clinical scores and could function as a clinical decision support tool to identify patients at high risk of PHLF pre- and postoperatively.

Distributed Maritime Operations and the Medical Challenge: How a Medical Common Operating Picture Can Help Bridge the Gap.

This work explores the challenges of delivering medical care in the geographically dispersed and resource-constrained environment of Distributed Maritime Operations (DMO) and Expeditionary Advanced Base Operations (EABO). Traditional medical planning approaches may struggle to adapt to the vast operational space, extended evacuation times, and limited medical force present in these scenarios. The concept of a Medical Common Operating Picture (COP) emerges as a potential solution. By providing a shared view of the medical situation across the theater, encompassing logistics, personnel, and patient data, a medical COP has the potential to facilitate medical command and control (MED C2) in DMO/EABO. The implementation of a medical COP has the potential to optimize resource allocation, enhance situational awareness, streamline medical evacuation, and reduce healthcare provider moral injury in large-scale combat operations. A medical COP will allow medical planners to make informed decisions on triage, resupply, and evacuation, ensuring the best use of limited medical resources. This is done by leveraging a comprehensive understanding of the medical landscape, enabling informed clinical and operational decision-making by humanitarian and combat personnel respectively. A fully realized medical COP system will enable a dynamic theater evacuation policy, balancing the conflicting needs of patient care at higher echelons with the operational expediency of returning servicemembers to their operational units, thereby maximizing evacuation effectiveness. It will further enable medical personnel to perform dynamic casualty triage based on operational realities, mitigating potential ethical dilemmas. Implementing such a medical COP system will require overcoming communication limitations to facilitate data exchange and potentially integrating clinical decision support tools for real-time data analysis and recommendations. It will also require the rapid adoption of modernized operational medicine documentation solutions by medical assets within the operational forces. Ultimately, this work suggests that a medical COP has the potential to bridge the gap between traditional medical planning and the unique demands of DMO/EABO, ultimately optimizing casualty care, maximizing resource efficiency, and preserving the fighting force.

550-P: Utilization of an Electronic Clinical Decision Support Tool by Health Care Providers for Hospital Glycemic Management in a Large Academic Center

550-P: Utilization of an Electronic Clinical Decision Support Tool by Health Care Providers for Hospital Glycemic Management in a Large Academic Center

“Following the data”: Perceptions of and willingness to use clinical decision support tools to inform HIV care among Haitian clinicians

BackgroundClinical decision support (CDS) tools can support HIV care, including through case tracking, treatment and medication monitoring, and promoting provider compliance with care guidelines. There has been limited research into the technical, organizational, and behavioral factors that impact perceptions of and willingness to use CDS tools at scale in resource-limited settings, including in Haiti. MethodsOur sample included fifteen purposively chosen Haitian HIV program experts, including active clinicians and HIV program managers. Participants completed structured quantitative surveys and one-on-one qualitative semi-structured interviews. ResultsStudy participants had high levels of familiarity and experience with CDS tools. The primary motivator for CDS tool use was a perceived benefit to quality of care, including improved provider time use, efficiency, and decision-making ability, and patient outcomes. Participants highlighted decision-making autonomy and how CDS tools could support provider decision making but should not supplant provider knowledge and experience. Participants highlighted the need for sufficient provider training/sensitization, inclusion of providers in the system design process, and prioritization of tool user-friendliness as key mechanisms to drive tool use and impact. Some participants noted that systemic issues, such as limited laboratory capacity, may reduce the usefulness of CDS alerts, particularly concerning differentiated care and priority viral load testing. ConclusionRespondents had largely positive perceptions of EMRs and CDS tools, particularly due to perceived improvements in quality of care. To improve tool use, stakeholders should prioritize tool user-friendliness and provider training. Addressing systemic health system issues is necessary to unlock the full potential of these tools.

Expanding access to veterinary clinical decision support in resource-limited settings: a scoping review of clinical decision support tools in medicine and antimicrobial stewardship.

Digital clinical decision support (CDS) tools are of growing importance in supporting healthcare professionals in understanding complex clinical problems and arriving at decisions that improve patient outcomes. CDS tools are also increasingly used to improve antimicrobial stewardship (AMS) practices in healthcare settings. However, far fewer CDS tools are available in lowerand middle-income countries (LMICs) and in animal health settings, where their use in improving diagnostic and treatment decision-making is likely to have the greatest impact. The aim of this study was to evaluate digital CDS tools designed as a direct aid to support diagnosis and/or treatment decisionmaking, by reviewing their scope, functions, methodologies, and quality. Recommendations for the development of veterinary CDS tools in LMICs are then provided. The review considered studies and reports published between January 2017 and October 2023 in the English language in peer-reviewed and gray literature. A total of 41 studies and reports detailing CDS tools were included in the final review, with 35 CDS tools designed for human healthcare settings and six tools for animal healthcare settings. Of the tools reviewed, the majority were deployed in high-income countries (80.5%). Support for AMS programs was a feature in 12 (29.3%) of the tools, with 10 tools in human healthcare settings. The capabilities of the CDS tools varied when reviewed against the GUIDES checklist. We recommend a methodological approach for the development of veterinary CDS tools in LMICs predicated on securing sufficient and sustainable funding. Employing a multidisciplinary development team is an important first step. Developing standalone CDS tools using Bayesian algorithms based on local expert knowledge will provide users with rapid and reliable access to quality guidance on diagnoses and treatments. Such tools are likely to contribute to improved disease management on farms and reduce inappropriate antimicrobial use, thus supporting AMS practices in areas of high need.

Predicting Glaucoma Surgical Outcomes Using Neural Networks and Machine Learning on Electronic Health Records.

To develop machine learning (ML) and deep learning (DL) models to predict glaucoma surgical outcomes, including postoperative intraocular pressure, use of ocular antihypertensive medications, and need for repeat surgery. We identified glaucoma surgeries performed at Stanford from 2013-2024, with two or more postoperative visits with intraocular pressure (IOP) measurement. Patient features were identified from the electronic health record (EHR), including demographics, prior diagnosis and procedure codes, medications and eye exam findings. Classical ML and DL models were developed to predict which glaucoma surgeries would result in surgical failure, defined as (1) IOP not reduced by more than 20% of preoperative baseline on two consecutive postoperative visits, (2) increased classes of glaucoma medications, and (3) need for additional glaucoma surgery or revision of original surgery. A total of 2398 glaucoma surgeries of 1571 patients were included, of which 1677 surgeries met failure criteria. Random forest performed best for prediction of overall surgical failure, with accuracy of 75.5% and area under the receiver operator curve (AUROC) of 76.7%, similar to the deep learning model (accuracy 75.5%, AUROC 76.6%). Across all models, prediction performance was better for IOP outcomes (AUROC 86%) than need for an additional surgery (AUROC 76%) or need for additional glaucoma medication (AUC 70%). ML and DL algorithms can predict glaucoma surgery outcomes using structured data inputs from EHRs. Models that predict outcomes of glaucoma surgery may one day provide the basis for clinical decision support tools supporting surgeons in personalizing glaucoma treatment plans.

Impact on biomarker documentation in community oncology by optimizing clinical decision support.

11182 Background: Timing of biomarker results is critical to informing the selection of appropriate precision medicine-based treatment options for patients. Prior studies have demonstrated that low rates of biomarker testing may be driven by poor documentation. Our clinical decision support (CDS) tool aims to provide community oncology clinicians with a clear picture of the ever-growing available treatments. Evaluation of CDS data revealed an excess of biomarker results documented as “unknown” thought to be related to results availability. Methods: During treatment selection in the CDS tool, providers answer various prompts related to the patient’s disease, staging, treatment intent, and biomarker results. Prompts are aligned with NCCN guidelines for the respective cancer type. Changes were implemented to ease providers’ documentation of biomarker results. For adjuvant (adjNSCLC) and metastatic NSCLC (mNSCLC), prompts were relocated to align with when biomarker results were likely to be available. For metastatic castrate resistant prostate cancer (mCRPC), some biomarkers are only needed after a patient has received prior docetaxel and hormone therapy, thus prompts were moved to that scenario. Re-prompting of biomarker status for subsequent therapy was also implemented if the initial response was “unknown.” We report changes in percentage of known biomarkers specifically in adjNSCLC, mNSCLC, and mCRPC. Descriptive statistics were used to describe the rates of known biomarkers, and a chi-square test was used to analyze differences from pre and post changes. Results: In this evaluation, 85,493 biomarkers were documented. In adjNSCLC and mNSCLC, biomarker prompt changes increased documentation of known biomarkers from 69% to 78% and 59% to 79% pre- to post-implementation. In adjNSCLC, ALK and PD-L1 recording increased by ≥10% (P&lt;0.00001). For mNSCLC, documentation increased for all known biomarkers (P&lt;0.00001), including ≥15% increases for 8 of 10 biomarkers. In mCRPC, known biomarker documentation improved from 34% pre-implementation to 61% post-implementation. BRCA1/2, MSI, MMR, TMB, and PSMA biomarker recording increased by ≥20% (P&lt;0.00001) after prompt changes. Selected biomarkers are shown in the table. Conclusions: Adjusting prompts in a CDS tool to align with the availability of biomarkers in clinical practice increased the documentation of known biomarker results for adjNSCLC, mNSCLC, and mCRPC. Based on these findings, similar changes are being implemented for other cancer types in the CDS tool. [Table: see text]

Enhancing Clinical Decision Support in Nephrology: Addressing Algorithmic Bias Through Artificial Intelligence Governance

There has been a steady rise in the use of clinical decision support (CDS) tools to guide Nephrology, as well as general clinical care. Through guidance set by federal agencies and concerns raised by clinical investigators, there has been an equal rise in understanding whether such tools exhibit algorithmic bias leading to unfairness. This has spurred the more fundamental question of whether sensitive variables such as race should be included in CDS tools. In order to properly answer this question, it is necessary to understand how algorithmic bias arises. We break down three sources of bias encountered when using electronic health record data to develop CDS tools: (1) use of proxy variables, (2) observability concerns and (3) underlying heterogeneity. We discuss how answering the question of whether to include sensitive variables like race often hinges more on qualitative considerations than on quantitative analysis, dependent on the function that the sensitive variable serves. Based on our experience with our own institution’s CDS governance group, we show how health system-based governance committees play a central role in guiding these difficult and important considerations. Ultimately, our goal is to foster a community practice of model development and governance teams that emphasizes consciousness about sensitive variables and prioritizes equity.

Generative AI enhanced with NCCN clinical practice guidelines for clinical decision support: A case study on bone cancer.

e13623 Background: Bone cancer is a complex and challenging disease to diagnose and treat in clinical practice. Recently, generative AI, especially large language models (LLMs), has demonstrated potential as a decision support tool for cancer. However, most implementations have overlooked the integration of available cancer guidelines, such as the NCCN Bone Cancer Guidelines, in fine-tuning the outputs of generative AI models. Incorporating these guidelines into LLMs presents an opportunity to harness the extensive clinical knowledge they contain and improve the decision-support capabilities of the model. Methods: In this study, the aim is to enhance the LLM with cancer clinical guidelines to enable accurate medical decisions and personalized treatment recommendations. Therefore, we introduce a novel method for incorporating the NCCN Bone Cancer Guidelines into LLMs using a Binary Decision Tree (BDT) approach. The approach involves constructing a BDT based on NCCN Bone Cancer Guidelines, where internal nodes represent decision points from the Guidelines, and leaf node signify final treatment suggestions. Then the LLM makes decision at each internal node, considering a given patient's characteristics, and guides toward a treatment recommendation in the leaf node. To assess the efficacy of Guideline-enhanced LLMs, an oncologist from our team created 11 hypothetical osteosarcoma patients’ medical progress notes. Each note contains their demographics, medical history, current illness, physical exams, diagnostic tests. We tested three LLMs in the implementation (GPT-4, GPT-3.5, and PaLM 2) and compared the LLM-generated treatment recommendations with the gold standard treatment across four runs with different random seeds (random seeds is a setting to control the LLM outputs). The results are reported as the average of four runs. The original LLMs are used as baseline methods for comparison. Results: The table below provides a comparison between the performance of original LLMs and those augmented with cancer guidelines for osteosarcoma treatment recommendations. We can observe that the PaLM 2 model demonstrated superior performance compared to its counterparts, underscoring the effectiveness of integrating cancer guidelines into LLMs for decision support. Conclusions: The clinical decision support capabilities of the LLMs are promising when enhanced by NCCN Bone Cancer Guidelines using our approach. To fully exhibit the potential of our proposed method as a clinical decision support tool, further investigation into other subtypes of bone cancer should be conducted in the future study. [Table: see text]

The future is now: Implementation of standard of care, preemptive pharmacogenomic testing in patients with gastrointestinal cancers.

1546 Background: Preemptive germline pharmacogenomic (PGx) testing provides insight into the safety of high-risk chemotherapy and effective use of supportive care medications in gastrointestinal (GI) cancers. Advances in clinical practice guidelines and clinician support have accelerated the appetite for PGx testing, particularly for drugs with severe adverse reactions. The objective of this study was to assess operational outcomes associated with end-to-end electronic health record (EHR) integration of standard of care, preemptive multigene PGx testing in patients with GI cancers. Methods: Patients in two UCHealth GI oncology clinics were tested for five PGx genes related to 5-fluorouracil (5-FU, DPYD), irinotecan ( UGT1A1), and supportive care therapies ( CYP2C19, CYP2D6, CYP2C9). The PGx test order was embedded in all GI oncology treatment plans. Patient samples were genotyped, and PGx results were returned as discrete data to the EHR. Automated clinical decision support (CDS) tools provided PGx-guided recommendations at the point of prescribing. We leveraged the EPIS process framework and evaluated operational outcomes such as number of patients tested, lab turnaround time, patients with actionable phenotypes, and number of CDS alerts and clinical actions. Results: The analysis included 89 patients who received results for DPYD/UGT1A1 and 83 patients who received results for supportive care genes. Among samples tested, DPYD/UGT1A1 and supportive care gene results were returned to the EHR in an average of 4.7 ± 2 days (range 1-8) and 17.3 ± 3 days (range 9-24), respectively. The number of patients with abnormal phenotypes, defined as poor metabolizers for UGT1A1 and any non-normal phenotypes for the other genes, were UGT1A1, n=10 (11.2%); DPYD, n=4 (4.5%); CYP2C19, n=43 (51.8%); CYP2D6, n=31 (41.3%; 8 uninterpretable results); and CYP2C9, n=19 (22.9%). This equated to 14/89 (15.7%) and 68/83 (81.9%) patients having an abnormal result for DPYD/UGT1A1 and at least one supportive care gene, respectively. CDS alerts fired for 32 drug-gene interactions for 25 distinct patients. Of the alerts, 7 (21.9%) were for 5-FU or irinotecan, which showed for 7/89 patients (7.9%); 25 (78.1%) were for supportive care medications, which showed for 23/83 patients (27.7%). There were 9 clinical actions taken in response to the PGx alerts, 7 (77.8%) for 5-FU or irinotecan and 2 (22.2%) for supportive care drugs. Appropriate clinical actions were taken for 100% of patients with 5-FU or irinotecan drug-gene interactions. Conclusions: End-to-end EHR integration of PGx testing and prescribing guidance facilitated clinician adoption of preemptive, standard of care multigene PGx testing, with test results driving important medication changes. Future studies will evaluate the impact of PGx testing on clinical outcomes and cost effectiveness metrics in this population.

Use of an AI-powered visualization platform to support precision clinical decision-making in patients with breast cancer undergoing neoadjuvant chemotherapy.

e12618 Background: Chemotherapy to reduce cancer burden and tumor size ("downstaging") is widely used in neoadjuvant treatment of early-stage breast cancer (ESBC) and can have a dramatic impact on selection of further treatment courses, including surgery. The resulting decision-making process between physicians and cancer patients requires detailed and effective communication of the patient's disease status, and clear presentation of this information in Tumor Board settings. Artificial intelligence (AI) -powered visualization platforms offer the potential to aid in evaluating breast tumor location and morphology, and accordingly optimize treatment. Additionally, these tools offer enhanced communication with patients regarding future treatment options, particularly serving as clinical decision support (CDS) tools. Methods: We assessed the performance of an AI platform that employs dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) to render three-dimensional (3D) visualizations of breast tumors and 5 additional chest tissues on an independent validation dataset comprise of 100 ESBC cases with ground-truth tumor labels vetted by 2 breast-specialized radiologists. The platform provides 3D depictions of the tumor and its multi-tissue context, as well as functionalities for quantifying tumor morphology, tumor-to-landmark structure distances, excision volumes, and approximate surgical margins. We subsequently evaluated the tool's utility in a real-world setting to understand how the tool impacts clinical decision-making and patient communication, specifically around surgical planning in the context of neoadjuvant chemotherapy (NACT). Results: The platform generated accurate 3D tumor maps, as evidenced by a median Dice score of 75.6% and a median Hausdorff distances of 15.1 mm relative to the radiologist validation set. The automatic longest tumor dimension calculations were linearly correlated with radiologist direct measurements (Pearson correlation coefficient = 0.78) and had a median absolute difference of 6.3 mm. We found strong overlap between the platform’s ground-truth convex hulls, as shown by a median Dice score of 81.5%. When used by a skilled breast-specialized surgeon, the tool enabled rapid and intuitive patient-physician communication of the location and extent of disease, and potential degrees of downstaging to breast-conserving surgery. Conclusions: Detailed 3D tumor and surrounding multi-tissue depictions offer both qualitative and quantitative comprehension of cancer topology and may aid in formulating an optimal treatment approach for breast cancer treatment. We show that detailed depictions enhance communication between physicians and patients and may aid in streamlining decision-making.

Predictive models as a clinical decision support tool for genetic cancer risk assessment among Mexican patients at risk of hereditary colorectal cancer.

10617 Background: Hereditary cancer syndromes, including Lynch syndrome (LS) with an increased risk of colorectal cancer (CRC), have prompted the development of clinical criteria and predictive models (PM) to identify at-risk individuals. However, these criteria have been validated with limited representation from the Latino population. Several factors may influence CRC penetrance, potentially leading to variations in PM performance across populations. Given that CRC ranks as the third most prevalent cancer worldwide, the increasing incidence of early-onset disease highlights the need for expanding Genetic Cancer Risk Assessment (GCRA) services. Methods: This prospective study included Mexican patients who were diagnosed with CRC and met NCCN criteria for GCRA. Participants were enrolled at two referral centers in Mexico, the Instituto Nacional de Ciencias Medicas y Nutricion Salvador Zubiran, Mexico City and at the Hospital Zambrano Hellion TecSalud, Nuevo Leon. Participants' carrier risk was estimated using the PREMM5 model. Amsterdam II and revised Bethesda criteria, immunohistochemistry (IHC) for DNA mismatch repair (MMR) proteins in tumor tissue and multigene panel testing results were reviewed. For this analysis, only pathogenic and likely pathogenic (P/LP) variants in probands were reported. Results: From April 2016 to October 2023, 194 patients were included, with a median age at diagnosis of 45.0 (range 17-82) years, and 49.5% were women. P/LP variants were identified in 37.6% (n=73). 27.5% (n=50) were carriers of P/LP variants associated with LS (29 MLH1, 13 MSH2, 6 MSH6, and 2 PMS2). 11.8% (n=23) were carriers of P/LP variants in other genes (7 APC, 4 ATM, 3 CHEK2, 2 BRCA1, 2 TP53, 2 STK11, 1 PALB2, 1 BRIP1, and 1 MUTYH). Overall, carriers of PVs were diagnosed with CRC at a younger age than non-carriers (42 vs 46 years, p=0.034). The Amsterdam II criteria proved to be accurate (Sen 55%, Spe 91%, PPV 0.69, NPV 0.85), while the Bethesda criteria (Sen 98%, Spe 18%, PPV 0.29, NPV 0.96) and IHC for MMR proteins (Sen 94%, Spe 67%, PPV 0.61, NPV 0.95) were found to be highly sensitive. The area under the ROC curve for PREMM5 was 0.82 with a mean score 28.5 (2.1 – 50) in patients with LS and 6.0 (0.9 – 50) in non-carriers. Carriers of PVs had a higher rate of multiple primary malignancies than non-carriers (41.0% vs 19.0%, p=0.0014), and self-reported family history of CRC (65.7% vs 24.8%, p&lt;0.001). Conclusions: Our results showed the PREMM5 model and IHC for MMR proteins effectively prioritized patients eligible for germline genetic testing in resource-limited settings. The performance of PREMM5 in this Mexican cohort was similar to that reported in the validation study of this model in other populations. Considering the spectrum of identified P/LP variants, the multigene panel test should be used in the comprehensive evaluation of Mexican patients with CRC.