Artificial intelligence models in prediction of response to cardiac resynchronization therapy: a systematic review.

"table 1 unsupervised artificial intelligence models author and date design n definition of the primary outcome best algorithm n of clusters results bivona et al (2022) retrospective cohort study 200 death (median follow-up of 4 years) gaussian mixture model 3 3 response clusters: 10% 40% and 66% deaths (overall death rate: 26%); cluster 2 vs cluster 1: hr 1 84 (95% ci 1 35-2 50; p < 0 001); cluster 3 vs cluster 1: hr 2 23 (95% ci 1 71-2 90; p < 0 001) cikes et al (2018) rct 1106 death from any cause or a non-fatal hf event (average follow-up of 2 3 years) multiple kernel learning and k-means clustering 4 phenogroups1 and 3 were associated with a substantially better treatment effect of crt-d on the primary outcome (hr 0 36; 95% ci 0 19-0 68; p < 0 001 and hr 0 35; 95% ci 0 19-0 64; p < 0 001) than observed in the other groups fenny et al (2020) retrospective cohort study 840 composite of death heart transplant placement of left ventricular assist device at 12-month follow-up k-means clustering 2 compared with group 2 group 1 had lower risk for reaching the composite end point (hr 0 44 [95% ci 0 38-0 53]; p < 0 001) gallard et al (2021) prospective cohort study 250 decrease in lvesv of >= 15% at 6-month follow-up k-means clustering 5 five clusters were identified with response rates of 50% 71% 72% 86% and 93% respectively (overall response rate was74%) gallard et al (2021) prospective cohort study 250 composite of death and hospitalization for hf (mean duration of follow-up was3 7 years) k-means clustering 5 five clusters were identified with adverse event rates of 37% 14% 14% 14% and 7% galli et al (2021) prospective cohort study 193 composite of heart transplantation lv assist device implantation or all-cause death (median follow-up of 37 months) k-medoid clustering 2 in comparison to group 1 group 2 had higher outcome reach: hr 4 70 (95% ci 2 1-10 0) galli et al (2021) prospective cohort study 193 decrease in lvesv of >= 15% at 6-month follow-up k-medoid clustering 2 groups1 and 2 had 89% and 34% responders respectively (overall response rate: 68%) riolet et al (2021) retrospective cohort study 328 death from any cause (median follow-up of 51 months) agglomerative hierarchical clustering based on k-means4 in comparison to group 1 groups2 3 and 4 had greater risk of death: hr 0 89 (95% ci 0 47-1 70) hr 3 23 (95% cl 1 9-5 5) and hr 2 49 (95% ci 1 38-4 50) riolet et al (2021) retrospective cohort study 328 decrease in lv end-systolic volume of >= 15% at 9-month follow-up agglomerative hierarchical clustering based on k-means4 in groups1 to 4 the response rates were 81% 78% 39% and 59% respectively hf heart failure lv left ventricle lvesv left ventricle end-systolic volume rct randomized controlled trial table 2 supervised artificial intelligence models author and date design n definition of the primary outcome best algorithm performance of the algorithm validation strategy (training/validation proportion) availability bivona et al (2022) retrospective cohort study 200 death (median follow-up of 4 years) logistic regression auc 0 86 internal fivefold cross-validation http://gmmxcrt pythonanywhere com cai et al (2021) retrospective cohort study 1664 absolute improvement of > 5% in lvef measured at 6-month follow-up stacked contractive autoencoder and ensemble of bagging/adaboost/xgboost auc 0 76 internal fivefold cross-validation (70%) internal hold-out test set (30%) nr feeny et al (2019) retrospective cohort study 925 10% absolute lvef increase at 12-month follow-up naive bayes classifier auc 0 70 internal fivefold cross-validation http://riskcalc org:3838/crtresponsescore fernandes et al (2023) prospective cohort study 158 lvef improvement of >= 5% at 6-month follow-up prediction analysis of microarrays auc 0 80 sensitivity 0 86 specificity 0 75 internal threefold cross-validation with 25 repeats internal hold-out test set (20%) nr field et al (2020) rct 419 decrease in lvesv >= 15% at 6-month follow-up classification tree probability of response in range from 30 to 63% internal cross-validation yes in the publication gallard et al (2020) retrospective cohort study 323 decrease in lvesv >= 15% at the 6-month follow-up random forest with feature selection auc 0 81 internal monte carlo cross-validation (80%/20%) nr galli et al (2021) prospective cohort study 193 decrease in lvesv >= 15% at 6-month follow-up ensemble learning: boruta algorithm and random forest auc 0 81 internal validation nr galli et al (2021) prospective cohort study 193 composite of heart transplantation lv assist device implantation or all-cause death (median follow-up of 37 months) ensemble learning: boruta algorithm and random forest auc 0 84 internal validation nr haque et al (2022) rct 794 decrease in esv >= 15 ml at 6 months post-implantation ensemble of nine equally weighted algorithms auc 0 78 internal fivefold cross-validation (45%) internal hold-out test set (20%) https://github com/sysmechbiolab/crt_iml he et al (2023) retrospective cohort study 130 lvef improvement of >= 5% at 6-month follow-up ensemble of unsupervised neural network and logistic regression auc 0 74 external validation nr hong et al (2022) prospective cohort study 280 adverse outcome (-) [death heart transplantation extracorporeal membrane oxygenation (ecmo) or use of a ventricular assist device] or lvef increase at 12-month follow-up or earlier classification tree - internal validation yes in the publication howell et al (2021) rct 741 freedom from death hf hospitalization and a decrease in lvesv >= 15% at 6-month follow-up adaptive lasso model auc 0 76 internal tenfold cross-validation (80%) internal hold-out test set (20%) https://ecgpredictscd org/crt hu et al (2019) retrospective cohort study 990 composite of < 0% improvement in lvef 6-18 months post-procedure or death by 18 months ensemble learning: machine learning and natural language processing auc 0 75 65% accuracy 79% specificity 26% sensitivity internal fivefold cross-validation (80%) internal hold-out test set (20%) nr kalscheur et al (2018) rct 1076 death or heart failure hospitalization at 12-month follow-up random forest auc 0 74 52% sensitivity 38% npv 80% specificity 88% ppv internal tenfold cross-validation (45%) internal hold-out test set (55%) nr lei et al (2019) retrospective cohort study 117 improvement in nyha functional >= 1 and a decrease in lvesv >= 15% at 6-month follow-up support vector machines85% accuracy internal tenfold cross-validation nr liang et al (2021) retrospective cohort study 752 > 10% absolute lvef increase at 12-month follow-up ridge regression auc 0 77 92% specificity 54% sensitivity internal tenfold cross-validation http://www crt-response com schmitz et al (2014) case-control study 156 decrease in lvesv >= 15% at 6- to 12-month follow-up (median 9 months) part algorithm 85% accuracy 88% specificity 82% sensitivity internal tenfold cross-validation yes in the publication tokodi et al (2020) retrospective cohort study 1510 death at 5-year follow-up random forest auc 0 80 internal tenfold cross-validation https://arguscognitive com/crt wouters et al (2023) retrospective cohort study 1306 left ventricular assist device (lvad) implantation heart transplantation (htx) and all-cause mortality (median follow-up 3 5 years) factorecg (neural network consisting from variational auto-encoder and decoder) auc 0 69 internal validation by means of boostrap https://crt ecgx ai wouters et al (2023) retrospective cohort study 821 decrease in lvesv >= 15% at 6 to 12-month follow-up factorecg (neural network consisting from variational auto-encoder and decoder) auc 0 69 internal validation by means of boostrap https://crt ecgx ai auc area under the receiver operating characteristic curve hf heart failure lv left ventricle lvef left ventricle ejection fraction lvesv left ventricle end-systolic volume npv negative predictive value rct randomized controlled trial ppv positive predictive value the risk of bias (rob) of included studies was performed by the first author of the study (wn)."

"Declarations Ethical approvalNot applicable. Competing interestsThe authors declare no competing interests. Competing interests The authors declare no competing interests."