TY - JOUR
T1 - Identifying subtypes of heart failure from three electronic health record sources with machine learning
T2 - an external, prognostic, and genetic validation study
AU - Banerjee, Amitava
AU - Dashtban, Ashkan
AU - Chen, Suliang
AU - Pasea, Laura
AU - Thygesen, Johan H.
AU - Fatemifar, Ghazaleh
AU - Tyl, Benoit
AU - Dyszynski, Tomasz
AU - Asselbergs, Folkert W.
AU - Lund, Lars H.
AU - Lumbers, Tom
AU - Denaxas, Spiros
AU - Hemingway, Harry
N1 - Funding Information: All authors are supported by the BigData@Heart Consortium, funded by the European Union Innovative Medicines Initiative-2 joint undertaking under grant agreement number 116074. This joint undertaking receives support from the EU's Horizon 2020 research and innovation programme and the European Federation of Pharmaceutical Industries and Associations. Publisher Copyright: © 2023 The Author(s). Published by Elsevier Ltd. This is an Open Access article under the CC BY 4.0 license
PY - 2023/6/1
Y1 - 2023/6/1
N2 - Background: Machine learning has been used to analyse heart failure subtypes, but not across large, distinct, population-based datasets, across the whole spectrum of causes and presentations, or with clinical and non-clinical validation by different machine learning methods. Using our published framework, we aimed to discover heart failure subtypes and validate them upon population representative data. Methods: In this external, prognostic, and genetic validation study we analysed individuals aged 30 years or older with incident heart failure from two population-based databases in the UK (Clinical Practice Research Datalink [CPRD] and The Health Improvement Network [THIN]) from 1998 to 2018. Pre-heart failure and post-heart failure factors (n=645) included demographic information, history, examination, blood laboratory values, and medications. We identified subtypes using four unsupervised machine learning methods (K-means, hierarchical, K-Medoids, and mixture model clustering) with 87 of 645 factors in each dataset. We evaluated subtypes for (1) external validity (across datasets); (2) prognostic validity (predictive accuracy for 1-year mortality); and (3) genetic validity (UK Biobank), association with polygenic risk score (PRS) for heart failure-related traits (n=11), and single nucleotide polymorphisms (n=12). Findings: We included 188 800, 124 262, and 9573 individuals with incident heart failure from CPRD, THIN, and UK Biobank, respectively, between Jan 1, 1998, and Jan 1, 2018. After identifying five clusters, we labelled heart failure subtypes as (1) early onset, (2) late onset, (3) atrial fibrillation related, (4) metabolic, and (5) cardiometabolic. In the external validity analysis, subtypes were similar across datasets (c-statistics: THIN model in CPRD ranged from 0·79 [subtype 3] to 0·94 [subtype 1], and CPRD model in THIN ranged from 0·79 [subtype 1] to 0·92 [subtypes 2 and 5]). In the prognostic validity analysis, 1-year all-cause mortality after heart failure diagnosis (subtype 1 0·20 [95% CI 0·14–0·25], subtype 2 0·46 [0·43–0·49], subtype 3 0·61 [0·57–0·64], subtype 4 0·11 [0·07–0·16], and subtype 5 0·37 [0·32–0·41]) differed across subtypes in CPRD and THIN data, as did risk of non-fatal cardiovascular diseases and all-cause hospitalisation. In the genetic validity analysis the atrial fibrillation-related subtype showed associations with the related PRS. Late onset and cardiometabolic subtypes were the most similar and strongly associated with PRS for hypertension, myocardial infarction, and obesity (p<0·0009). We developed a prototype app for routine clinical use, which could enable evaluation of effectiveness and cost-effectiveness. Interpretation: Across four methods and three datasets, including genetic data, in the largest study of incident heart failure to date, we identified five machine learning-informed subtypes, which might inform aetiological research, clinical risk prediction, and the design of heart failure trials. Funding: European Union Innovative Medicines Initiative-2.
AB - Background: Machine learning has been used to analyse heart failure subtypes, but not across large, distinct, population-based datasets, across the whole spectrum of causes and presentations, or with clinical and non-clinical validation by different machine learning methods. Using our published framework, we aimed to discover heart failure subtypes and validate them upon population representative data. Methods: In this external, prognostic, and genetic validation study we analysed individuals aged 30 years or older with incident heart failure from two population-based databases in the UK (Clinical Practice Research Datalink [CPRD] and The Health Improvement Network [THIN]) from 1998 to 2018. Pre-heart failure and post-heart failure factors (n=645) included demographic information, history, examination, blood laboratory values, and medications. We identified subtypes using four unsupervised machine learning methods (K-means, hierarchical, K-Medoids, and mixture model clustering) with 87 of 645 factors in each dataset. We evaluated subtypes for (1) external validity (across datasets); (2) prognostic validity (predictive accuracy for 1-year mortality); and (3) genetic validity (UK Biobank), association with polygenic risk score (PRS) for heart failure-related traits (n=11), and single nucleotide polymorphisms (n=12). Findings: We included 188 800, 124 262, and 9573 individuals with incident heart failure from CPRD, THIN, and UK Biobank, respectively, between Jan 1, 1998, and Jan 1, 2018. After identifying five clusters, we labelled heart failure subtypes as (1) early onset, (2) late onset, (3) atrial fibrillation related, (4) metabolic, and (5) cardiometabolic. In the external validity analysis, subtypes were similar across datasets (c-statistics: THIN model in CPRD ranged from 0·79 [subtype 3] to 0·94 [subtype 1], and CPRD model in THIN ranged from 0·79 [subtype 1] to 0·92 [subtypes 2 and 5]). In the prognostic validity analysis, 1-year all-cause mortality after heart failure diagnosis (subtype 1 0·20 [95% CI 0·14–0·25], subtype 2 0·46 [0·43–0·49], subtype 3 0·61 [0·57–0·64], subtype 4 0·11 [0·07–0·16], and subtype 5 0·37 [0·32–0·41]) differed across subtypes in CPRD and THIN data, as did risk of non-fatal cardiovascular diseases and all-cause hospitalisation. In the genetic validity analysis the atrial fibrillation-related subtype showed associations with the related PRS. Late onset and cardiometabolic subtypes were the most similar and strongly associated with PRS for hypertension, myocardial infarction, and obesity (p<0·0009). We developed a prototype app for routine clinical use, which could enable evaluation of effectiveness and cost-effectiveness. Interpretation: Across four methods and three datasets, including genetic data, in the largest study of incident heart failure to date, we identified five machine learning-informed subtypes, which might inform aetiological research, clinical risk prediction, and the design of heart failure trials. Funding: European Union Innovative Medicines Initiative-2.
UR - http://www.scopus.com/inward/record.url?scp=85160019705&partnerID=8YFLogxK
U2 - https://doi.org/10.1016/S2589-7500(23)00065-1
DO - https://doi.org/10.1016/S2589-7500(23)00065-1
M3 - Article
C2 - 37236697
SN - 2589-7500
VL - 5
SP - e370-e379
JO - The Lancet Digital Health
JF - The Lancet Digital Health
IS - 6
ER -