@article{fdi:010083269,
  title = {{D}eep learning analysis of electrocardiogram for risk prediction of drug-induced arrhythmias and diagnosis of long {QT} syndrome},
  author = {{P}rifti, {E}di and {F}all, {A}. and {D}avogustto, {G}. and {P}ulini, {A}. and {D}enjoy, {I}. and {F}unck-{B}rentano, {C}. and {K}han, {Y}. and {D}urand-{S}almon, {A}. and {B}adilini, {F}. and {W}ells, {Q}. {S}. and {L}eenhardt, {A}. and {Z}ucker, {J}ean-{D}aniel and {R}oden, {D}. {M}. and {E}xtramiana, {F}. and {S}alem, {J}. {E}.},
  editor = {},
  language = {{ENG}},
  abstract = {{A}ims {C}ongenital long-{QT} syndromes (c{LQTS}) or drug-induced long-{QT} syndromes (di{LQTS}) can cause torsade de pointer ({T}d{P}), a life-threatening ventricular arrhythmia. {T}he current strategy for the identification of drugs at the high risk of {T}d{P} relies on measuring the {QT} interval corrected for heart rate ({QT}c) on the electrocardiogram ({ECG}). {H}owever, {QT}c has a low positive predictive value. {M}ethods and results {W}e used convolutional neural network ({CNN}) models to quantify {ECG} alterations induced by sotalol, an {IK}r blocker associated with {T}d{P}, aiming to provide new tools ({CNN} models) to enhance the prediction of drug-induced {T}d{P} (di{T}d{P}) and diagnosis of c{LQTS}. {T}ested {CNN} models used single or multiple 10-s recordings/patient using 8 leads or single leads in various cohorts: 1029 healthy subjects before and after sotalol intake (n = 14 135 {ECG}s); 487 c{LQTS} patients (n = 1083 {ECG}s: 560 type 1, 456 type 2, 67 type 3); and 48 patients with di{T}d{P} (n = 1105 {ECG}s, with 147 obtained within 48 h of a di{T}d{P} episode). {CNN} models outperformed models using {QT}c to identify exposure to sotalol [area under the receiver operating characteristic curve ({ROC}-{AUC}) = 0.98 vs. 0.72, {P} <= 0.001]. {CNN} models had higher {ROC}-{AUC} using multiple vs. single 10-s {ECG} ({P} <= 0.001). {P}erformances were comparable for 8-lead vs. single-lead models. {CNN} models predicting sotalol exposure also accurately detected the presence and type of c{LQTS} vs. healthy controls, particularly for c{LQT}2 ({AUC}-{ROC} = 0.9) and were greatest shortly after a di{T}d{P} event and declining over time ({P} <= 0.001), after controlling for {QT}c and intake of culprit drugs. {ECG} segment analysis identified the {J}-{T}peak interval as the best discriminator of sotalol intake. {C}onclusion {CNN} models applied to {ECG}s outperform {QT}c measurements to identify exposure to drugs altering the {QT} interval, congenital {LQTS}, and are greatest shortly after a di{T}d{P} episode. [{GRAPHICS}] .},
  keywords = {{T}orsades de pointes ; {M}achine learning ; {R}isk prediction ; {I}nterpretability ; {L}ong {QT}},
  booktitle = {},
  journal = {{E}uropean {H}eart {J}ournal},
  volume = {42},
  numero = {38},
  pages = {3948--3961},
  ISSN = {0195-668{X}},
  year = {2021},
  DOI = {10.1093/eurheartj/ehab588},
  URL = {https://www.documentation.ird.fr/hor/fdi:010083269},
}