Machine learning algorithms using large datasets may be utilized to accurately estimate prognosis and guide therapy for patients with adult congenital heart disease, according to a study published in the European Heart Journal.

The investigators of this large cohort, single-center study sought to examine the utility of machine learning algorithms as a prognostic model and to guide therapeutic decision-making in patients with adult congenital heart disease or pulmonary hypertension.

The study sample included 10,019 adults under active follow-up at the Royal Brompton Hospital in London between 2000 and 2018. Patient data were retrospectively collected — including clinical and demographic data, ECG parameters, cardiopulmonary exercise data, and laboratory markers — and incorporated into deep learning algorithms. Specific deep learning models were then built for patient categorization into diagnostic subsets, disease complexity subsets, and by New York Heart Association (NYHA) class. Accuracy was calculated for all deep learning models, which were used to gauge the prognosis of individual patients as well as to estimate the need for discussion with a multidisciplinary team.

Overall, the deep learning models were able to accurately categorize diagnosis in 91.1% of the study sample, disease complexity in 97% of the sample, and NYHA class in 90.6% of the sample. The need for patient multidisciplinary team discussion was accurately predicted in 90.2% of the sample. Based on a median follow-up time of 8 years, in which 785 patients died, the investigators constructed a model to estimate mortality based on age, diagnosis, laboratory data, ECG parameters, and cardiopulmonary exercise data. Independent of demographic, exercise, laboratory, and ECG parameters, univariate Cox analysis showed a disease severity score greater than 0.9 was related to all-cause mortality with a hazard ratio of 34.02 (<.0001).

Limitations to the study included a structured letter format required for the text mining algorithm, limiting the number of laboratory and ECG parameters analyzed, and the need for specifically trained networks for future incorporation of imaging data. Another limitation was the single-center study design, which may not be representative of adult congenital heart disease patterns in the general population. Finally, the generalizability of the model is limited as the text analysis is language specific and requires adaptation for use in different language settings.   

The investigators suggest that machine learning algorithms using large text-driven datasets could predict individual patient’s prognosis with accuracy comparable to manual coding and that utilizing these models may provide clinical guidance for improving the care of patients with adult congenital heart disease.

This study was sponsored by the British Heart Foundation, Actelion UK, Pfizer UK, and GSK UK.

Reference

Diller GP, Kempny A, Babu-Narayan SV, et al. Machine learning algorithms estimating prognosis and guiding therapy in adult congenital heart disease: data from a single tertiary center including 10,019 patients [published online January 26, 2019]. Eur Heart J. doi:10.1093/eurheartj/ehy915