TY - JOUR T1 - Territory-wide cohort study of Brugada syndrome in Hong Kong: predictors of long-term outcomes using random survival forests and non-negative matrix factorisation JF - Open Heart JO - Open Heart DO - 10.1136/openhrt-2020-001505 VL - 8 IS - 1 SP - e001505 AU - Sharen Lee AU - Jiandong Zhou AU - Ka Hou Christien Li AU - Keith Sai Kit Leung AU - Ishan Lakhani AU - Tong Liu AU - Ian Chi Kei Wong AU - Ngai Shing Mok AU - Chloe Mak AU - Kamalan Jeevaratnam AU - Qingpeng Zhang AU - Gary Tse Y1 - 2021/02/01 UR - http://openheart.bmj.com/content/8/1/e001505.abstract N2 - Objectives Brugada syndrome (BrS) is an ion channelopathy that predisposes affected patients to spontaneous ventricular tachycardia/fibrillation (VT/VF) and sudden cardiac death. The aim of this study is to examine the predictive factors of spontaneous VT/VF.Methods This was a territory-wide retrospective cohort study of patients diagnosed with BrS between 1997 and 2019. The primary outcome was spontaneous VT/VF. Cox regression was used to identify significant risk predictors. Non-linear interactions between variables (latent patterns) were extracted using non-negative matrix factorisation (NMF) and used as inputs into the random survival forest (RSF) model.Results This study included 516 consecutive BrS patients (mean age of initial presentation=50±16 years, male=92%) with a median follow-up of 86 (IQR: 45–118) months. The cohort was divided into subgroups based on initial disease manifestation: asymptomatic (n=314), syncope (n=159) or VT/VF (n=41). Annualised event rates per person-year were 1.70%, 0.05% and 0.01% for the VT/VF, syncope and asymptomatic subgroups, respectively. Multivariate Cox regression analysis revealed initial presentation of VT/VF (HR=24.0, 95% CI=1.21 to 479, p=0.037) and SD of P-wave duration (HR=1.07, 95% CI=1.00 to 1.13, p=0.044) were significant predictors. The NMF-RSF showed the best predictive performance compared with RSF and Cox regression models (precision: 0.87 vs 0.83 vs. 0.76, recall: 0.89 vs. 0.85 vs 0.73, F1-score: 0.88 vs 0.84 vs 0.74).Conclusions Clinical history, electrocardiographic markers and investigation results provide important information for risk stratification. Machine learning techniques using NMF and RSF significantly improves overall risk stratification performance. ER -