Table 1.
The incidence of postoperative complication (prediction variables) for the three datasets before the split into validation/training sets are depicted above with the number of patients experiencing each variable labelled.
PD: Pancreaticoduodenectomy.
Fig 1.
5.881.881 patients were in the National Surgical Quality Improvement Program (NSQIP) dataset, 31.944 of whom were PD patients.
216 of these patients were excluded because of an operation time of less than 120 minutes. The remaining 31.728 patients were split into two datasets. One dataset with 40% of the PD patients which was recombined the with the patients from the remainder of the NSQIP dataset (General dataset and the second data frame which was the 60% were split into a training set, validation set and a test set that was used after the training of all the models to test their accuracy.
Fig 2.
Model architecture with all layers depicted.
Fig 3.
Performance metrics measures as Receiver Operator Characteristics Area Under the Curve (ROC AUC) of the four different modelling approaches benchmarked against each other for predicting the mortality and the 18 different complications included in the National Surgical Quality Improvement Program (NSQIP) dataset.
SSSI: Superficial Surgical site infection, DSSI: Deep surgical site infection, OSSI: Organ/space surgical site infection, WOUND: Wound disruption, PNEUMONIA: Postoperative pneumonia, UNPINT: Unplanned intubation, PE: Pulmonary embolism, VENT48: Ventilator dependence >48 hours, PRI: Progressive renal insufficiency, ARF: Acute renal failure, UTI: Urinary tract infection, STROKE: Stroke, CAR: Cardiac arrest requiring CPR, MI: Myocardial infarction, DVT: Deep vein thrombosis, SEPSIS: Sepsis, SEPSHOCK: Septic shock, BLEED: Bleeding requiring transfusion, DECEASED: Mortality.
Table 2.
The overall performance of the four models on all variables in the test set with Receiver Operator Characteristics Area Under the Curve (ROC_AUC) values as the metric.
The general model was trained on a general surgery patient cohort, the transfer learning model was trained on the general surgery patient cohort and transferred to a PD-specific patient cohort, the direct model, and the Random Forest model (RF) was trained exclusively on the PD-specific patient cohort.
Fig 4.
The average morbidity score and mortality Area Under the Receiver Operator Curve (ROC AUC) of the 4 models as well as the American College of Surgeons Surgical Risk Calculator (ACS-SRC).
Table 3.
The average morbidity Receiver Operator Characteristics Area Under the Curve (ROC_AUC) scores of the four models, calculated on the test set.
Additionally, the table includes the average morbidity and mortality risk scores obtained from the same test set derived from the American College of Surgeons Surgical Risk Calculator (ACS-SRC).
Fig 5.
SHAP values for the transfer learning model.
The x-axis contains the average impact on the color-coded prediction tasks and the y-axis represent the input variables hierarchically dependent on impact level. PATOS: Present at time of surgery.
Fig 6.
SHAP values for the direct model.
The x-axis contains the average impact on the color-coded prediction tasks and the y-axis represent the input variables hierarchically dependent on impact level. PATOS: Present at time of surgery.