TY - JOUR
T1 - Corrigendum
T2 - “Predicting respiratory failure after pulmonary lobectomy using machine learning techniques” (Surgery (2020) 168(4) (743–752), (S0039606020303317), (10.1016/j.surg.2020.05.032))
AU - Bolourani, Siavash
AU - Wang, Ping
AU - Patel, Vihas M.
AU - Manetta, Frank
AU - Lee, Paul C.
N1 - Publisher Copyright:
© 2020 Elsevier Inc.
PY - 2021/4
Y1 - 2021/4
N2 - We discovered an error in a line of the code used in the experimentation of the 2 methods presented in Fig 3.1 The error is in the section where the prediction models are being performed and introduces a “leak” between the training set and testing set. This error exists for both models presented. The implication of this error is that the predictive ability of the 2 models is overestimated. After correcting the error, the area under the receiver operating characteristic of the models diminishes by ~10%. The sensitivity and specificity of the model presented in Fig 3, A (random forest alone) are 40.1% and 96.6%, respectively. The sensitivity and specificity of the model in Fig 3, B (random forest +Combined over and under-sampling using SMOTE and Edited Nearest Neighbours [SMOTEENN] algorithm) are 69.4% and 85.0%, respectively. The discussion regarding Fig 3 remains the same: the random forest model was the accurate model; after applying the SMOTEENN method, the sensitivity of the model increases, which comes at the cost of decreasing specificity. The oversampling SMOTEENN model used is from Python library imblearn.combine.SMOTEENN. This error does not affect the methodology, the rest of the results, the discussion, or the conclusion of this article. For the reference, we provide the corrected version of the model presented in Fig 3. We would like to convey our sincere apologies to the readers and editors for the inconvenience this has caused.
AB - We discovered an error in a line of the code used in the experimentation of the 2 methods presented in Fig 3.1 The error is in the section where the prediction models are being performed and introduces a “leak” between the training set and testing set. This error exists for both models presented. The implication of this error is that the predictive ability of the 2 models is overestimated. After correcting the error, the area under the receiver operating characteristic of the models diminishes by ~10%. The sensitivity and specificity of the model presented in Fig 3, A (random forest alone) are 40.1% and 96.6%, respectively. The sensitivity and specificity of the model in Fig 3, B (random forest +Combined over and under-sampling using SMOTE and Edited Nearest Neighbours [SMOTEENN] algorithm) are 69.4% and 85.0%, respectively. The discussion regarding Fig 3 remains the same: the random forest model was the accurate model; after applying the SMOTEENN method, the sensitivity of the model increases, which comes at the cost of decreasing specificity. The oversampling SMOTEENN model used is from Python library imblearn.combine.SMOTEENN. This error does not affect the methodology, the rest of the results, the discussion, or the conclusion of this article. For the reference, we provide the corrected version of the model presented in Fig 3. We would like to convey our sincere apologies to the readers and editors for the inconvenience this has caused.
UR - http://www.scopus.com/inward/record.url?scp=85097452710&partnerID=8YFLogxK
U2 - 10.1016/j.surg.2020.10.031
DO - 10.1016/j.surg.2020.10.031
M3 - Comment/debate
C2 - 33288213
AN - SCOPUS:85097452710
SN - 0039-6060
VL - 169
SP - 1001
JO - Surgery (United States)
JF - Surgery (United States)
IS - 4
ER -