P3426Machine learning for medical decision support in a first attendance ambulatory of a tertiary care cardiologic hospital

• Published in: European heart journal Vol. 40; no. Supplement_1
• Main Authors: Hortegal, R, Babolin, B, Hsu, G J, Polivanov, R, Salvador, P, Brito, S, Santos, I, Regis, C D, Fragata, A, Szewierenko, P
• Format: Journal Article
• Language: English
• Published: Oxford University Press 01.10.2019
• Subjects: Prevention - Cardiovascular Risk Assessment: Scores
• ISSN: 0195-668X; 1522-9645
• DOI: 10.1093/eurheartj/ehz745.0300

Abstract Introduction Cardiovascular disease is an expensive public health problem. Establish the right level of healthcare attention for each patient in a high-demand system is a complex task, and in this scenario, the development of computational methods to support medical decisions has shown to be quite promising. Purpose Define Machine Learning (ML) algorithms to support medical decisions in a first attendance ambulatory of a tertiary cardiology hospital. Methods A prospective observational study was performed in 336 patients (58±13 years and 49.4% male), obtaining clinical and ECG/VCG data. A follow up of 15 months was performed in order to access MACE, PCI, Cardiac Surgery and evidence of Severe Cardiac Disease. From twenty-five initial features, running the ML K-means Clustering algorithm, we identify which ones to use and the optimal number of Clusters. Once defined the Clusters the data were labeled, and then the clusters compared with field data (outcomes) and by Kaplan Meyer curves. The labeled data were also run by a Gradient Boosting algorithm in order to define a Predictor for future use in medical decision support. Results The best result, with well-defined Clusters, was obtained with the combination 5 Clusters and 8 specific Features, and the follow-up data has matched the Cluster classification as shown in the Table. Kaplan Meyer curves corroborated these finding with statistically significant differences between the Clusters: Log-rank test (p<0.001). Predictor algorithm, trained by the labeled data, presented an average precision of 95% (CI 95%; 91–100%). Clustering Outcomes vs Follow-up results Cluster Patients Features Follow-up 15 months results Age BMI Previous Cardiac Previous Previous Diabetes** SM QRS_T QRS_T_loop Severe Heart Outcome (year)* (kg/m2)* Surgery** MI** PCI** angle* index* Disease* (%) 1 96 46±12 28.3±5.4 0 0.03 0.02 0 60±34 0.31±0.61 0.26 4.2% 2 60 61±12 29.6±5.5 0 0.25 0 1 97±45 0.41±1.32 0.57 20.0% 3 34 64±11 27.6±3.8 1 0.59 0.35 0.5 118±43 -0.11±1.22 0.72 35.3% 4 114 64±9 25.4±4.0 0 0.21 0 0 101±41 0.48±0.98 0.38 12.3% 5 32 50±10 28.7±5.1 0 0.88 1 0.47 82±34 0.52±0.69 0.81 50.0% *Mean ± sd; **yes = 1; no = 0. Conclusion The defined Predictor, using eight simple, quick and easy to get Features (clinical and ECG/VCG), shows excellent performance to classify patients who require tertiary cardiovascular healthcare attention.