5118
Machine-learning characterization of myocardial deformation patterns to identify responders to resynchronization therapy

Introduction: Heart failure (HF) is a heterogeneous syndrome and patients with distinct phenotypic characteristics may benefit from specific therapies. Cardiac resynchronization therapy (CRT), for example, may be ineffective in up to a third of patients, emphasizing the need for improved selection. We evaluated an unsupervised machine learning algorithm using myocardial deformation curves for predicting response to CRT.

Methods: In 1000 HF patients (LVEF≤30%, QRS≥130 ms) randomized to either CRT-D (n=596) or ICD (n=404) enrolled in MADIT-CRT we assessed strain patterns at 49 left ventricular locations from an apical view over a cardiac cycle using an unsupervised machine learning algorithm (multiple kernel learning) to position subjects based on similarities in deformation. A K-means algorithm identified the most plausible clustering configuration. We compared baseline characteristics, outcomes and treatment response for the primary outcome of death or heart failure event between clusters.

Results: The unsupervised analysis identified 4 clinically-distinct clusters (Figure) that significantly differed in age, gender, NYHA class, prevalence of ischemic heart disease and LBBB, QRS duration, LV size, EF, and BNP (Table). One such cluster (Cluster 2), comprising the highest proportion of known clinical characteristics predictive of CRT response, was associated with a substantially better treatment effect (HR 0.32, 95% CI 0.16, 0.65) than that observed in the overall cohort (interaction p=0.05).

	Cluster 1	Cluster 2	Cluster 3	Cluster 4	P value
	(n=151)	(n=275)	(n=390)	(n=184)
Age (years)	66.57±9.31	63.00±11.8	64.44±10.35	65.65±11.15	p=0.005
Female	41 (27.2%)	98 (35.6%)	71 (18.2%)	36 (19.6%)	p<0.001
NYHA II	129 (85.4%)	249 (90.5%)	317 (81.3%)	149 (81.0%)	p=0.006
QRS duration (ms)	150.78±14.76	162.43±18.16	157.98±20.01	152.62±18.72	p<0.001
LVEF (%)	26.87±3.82	22.85±5.45	23.48±5.17	25.20±4.64	p<0.001
Ischaemic CM	69 (45.7%)	108 (39.3%)	285 (73.1%)	109 (59.2%)	p<0.001
LBBB	99 (65.6%)	241 (87.6%)	250 (64.1%)	108 (59.0%)	p<0.001
LVEDV (mL)	216.75±36.97	254.58±66.09	258.19±59.76	234.66±38.00	p<0.001
BNP (pg/mL)	78.70±97.40	132.81±179.84	144.32±154.95	139.56±184.60	p=0.002

Conclusion: Our results serve as a proof-of-concept that machine-learning based approaches can identify clinically distinct patterns from deformation imaging in a phenotypically heterogeneous cohort and might aid in better recognition of disease-specific patterns that might benefit from specific therapies.

Acknowledgement/Funding: The MADIT-CRT trial was sponsored by Boston Scientific