Abstract:Background Despite evidence-based pharmacological recommendations, many patients with HFrEF remain undertreated. (1) These patients remain at risk of decompensation and might benefit treatment optimisation. Many patients may also have been diagnosed with HFrEF prior to the use of sacubitril/valsartan or SGLT2 inhibitors. Identification of patients with historical HF diagnosis is time-consuming. Artificial intelligence approaches using deep learning might allow rapid identification of such patients. Aim To develop an artificial intelligence deep-learning algorithm for rapid identification of HF from electronic health records (EHRs) and to determine the potential benefit of our AI-driven algorithm for the detection of undertreated HFrEF patients in the community. Methodology In this study data was collected from EHRs of 1,200 patients who underwent transthoracic echocardiography. The records were screened to identify HF diagnosis through an AI deep-learning algorithm using a Convolutional Neural Network (CNN). Medication status of patients identified as having HFrEF was assessed to determine whether they were on optimal guideline-directed medical therapy. Accuracy of the AI algorithm was assessed by a manual clinician review in a subset of 150 patients. Eligible patients not on optimal medical therapy where no further optimisation was planned (judged to be "stable" at last visit) were invited for a clinical evaluation including NT-proBNP and Kansas City Cardiomyopathy Questionnaire (KCCQ) and offered a treatment optimization that could include either uptitration or initiation of a new medication if symptomatic. NT-proBNP and KCCQ were repeated at 12 week, pre- and post-optimization results were compared to evaluate patient outcomes. Results The deep-learning algorithm accurately identified HFrEF patients (99%) and use of RAAS inhibitors (97%) and beta-blockers (92%). 73 patients attended for initial clinical assessment (mean age 70 years, 77% male) and demonstrated a moderate level of HF-symptom burden (median NT-proBNP was 679 pg/mL, interquartile range was from 220 to 1300) and mean KCCQ overall summary score was 70, indicating a moderate level of HF symptom burden. 27 patients agreed to treatment optimisation. After 12 weeks there was a significant 22% reduction in NT-proBNP (from 1991 pg/mL (SD+/- 1796) to 1630 pg/mL (SD +/- 1588),p=0.049) and improvements in KCCQ (total symptom score 79 (+/-22) to 85 (+/-20), p=0.005; clinical summary score 78 (+/-23) to 82 (+/-20); overall summary score from 75 (+/- 24) to 80 (+/-22), both p=0.04). 23 patients (85%) and 18 patients (67%) out of 27 had an improvement in NT-proBNP and KCCQ overall summary scores respectively. Conclusion In our study we demonstrated the feasibility and potential benefits incorporating AI into a clinical workflow, allowing optimising treatment of patients with HF, with improvements in biomarkers and quality of life that could translate to long-term patient benefits.Changes in NT-proBNP after treatmentChanges in KCCQ after treatment

Development, validation and prospective clinical implementation of a machine learning algorithm for incident cardio-renal-metabolic diseases and cardiovascular death: the OPTIMISE study

Machine learning identifies individuals at higher risk of incident cardio-renal-metabolic diseases and cardiovascular death who have unrealised opportunities to reduce future cardiovascular risk

Improving Cardiovascular Risk Prediction Through Machine Learning Modelling of Irregularly Repeated Electronic Health Records

Cardiovascular Event Prediction by Machine Learning: The Multi-Ethnic Study of Atherosclerosis

Improving 10-year cardiovascular risk prediction in apparently healthy people: flexible addition of risk modifiers on top of SCORE2

Integrating artificial intelligence into a real-world clinical pathway to facilitate clinician treatment optimisation in patients with HFrEF on suboptimal medical therapy

A Web-Based Application for Risk Stratification and Optimization in Patients With Cardiovascular Disease: Pilot Study

Cardiovascular risk prediction in healthy older people

Machine Learning For Risk Prediction After Heart Failure Emergency Department Visit or Hospital Admission Using Administrative Health Data

Predicting the risks of kidney failure and death in adults with moderate to severe chronic kidney disease: multinational, longitudinal, population based, cohort study

Cardio-pulmonary exercise testing in adult congenital heart patients: a new data-focused approach to predict mortality

Machine Learning Improves Upon Clinicians' Prediction of End Stage Kidney Disease

Cardiovascular risk prediction using metabolomic biomarkers and polygenic risk scores: a cohort study and modelling analyses

#5650 METABOLIC BLOOD BIOMARKER PROFILING FOR CHRONIC KIDNEY DISEASE PREDICTION – EVIDENCE FROM 275,000 INDIVIDUALS IN THE UK BIOBANK

Proactive Identification of Patients with Diabetes at Risk of Uncontrolled Outcomes during a Diabetes Management Program: Conceptualization and Development Study Using Machine Learning

Accurate prediction of all-cause mortality in patients with metabolic dysfunction-associated steatotic liver disease using electronic health records

Predicting incident heart failure from population-based nationwide electronic health records: protocol for a model development and validation study

Clinical evaluation of a machine learning-based early warning system for patient deterioration

Computational and Human Intelligence Methods for Constructing Practical Risk Prediction Models: An Application to Cardio-Renal Outcomes in Non-Diabetic CKD Patients

Prognosis and Personalized In Silico Prediction of Treatment Efficacy in Cardiovascular and Chronic Kidney Disease: A Proof-of-Concept Study