Prediction Of Short-term Postoperative Complications In Gastrointestinal Cancer Patients Based On Artificial Intelligence

PurposeThe purpose of this study was to screen the risk factors associated with short-term complications after radical operation for gastrointestinal cancer,and construct nomogram models to predict the high-risk population for complications,and further verify the prediction efficiency.MethodsA total of 335 patients who received radical operation for gastric cancer and colorectal cancer in Yuhuangding Hospital of Yantai Affiliated to Qingdao University from June 2020 to July 2021 were included in this study.They were divided into a training set（268 cases）and a verification set（67 cases）according to the time sequence.All patients strictly monitored the perioperative routine blood test,biochemical test and other indicators and recorded the postoperative fluid access.They closely followed the occurrence of short-term postoperative complications,and conducted early evaluation and standardized treatment of the complications.One hundred and sixty-eight patients were monitored for perioperative stress indicators,including C-reactive protein（CRP）,interleukin-6（IL-6）and cortisol,and the correlation between postoperative complications and stress changes as well as changes in fluid intake and output was explored.Based on the results of routine test indexes of the patients in the training set and the liquid inlet and outlet variables verified and determined above,single-factor and multi-factor Logistics regression analysis and machine learning algorithms in artificial intelligence are used,including random forest（RF）,support vector machine（SVM）,artificial neural network（ANN）,respectively,to determine the risk factors associated with complications occurring within 7 days after surgery and complications occurring within 30 days after surgery.Furthermore,based on the identified risk factors and the principle of nomogram construction,the nomogram prediction models of complications within 7 days and 30 days after surgery based on Logistics regression and artificial intelligence were established,respectively.Based on the data of the training set and the verification set,the receiver operating characteristic curve（ROC curve）of the two models was constructed,and the prediction accuracy of the two models was compared according to the area under the curve–—AUC value.The calibration curve was finally constructed to evaluate the accuracy and effectiveness of the model.Results1.The 335 patients were divided into a training set（268 cases）and a verification set（67 cases）according to the time sequence.There was no statistical difference between groups in the baseline characteristics such as age,gender,BMI,and diabetes,short-term outcomes such as postoperative hospital stay,fasting time,and first exhaust time,and pathological outcomes such as lymph node metastasis,vascular cancer embolism,and nerve infiltration（all P>0.05）.2.A total of 86 cases of complications within 7 days after surgery and 102 cases of complications within 30 days（including 7-day complications）after surgery were counted,including 81 cases in the training set and 21 cases in the verification set.There was no significant difference between the two groups（P=0.859）.There were 20 severe complications（Clavien-dindo score III/IV/V）,including 16 cases in the training set and4 cases in the verification set.There was no significant difference between the two groups（P=1.000）.3.There was no significant difference in the preoperative levels of CRP,IL-6 and cortisol between patients with complications and without complications（all P>0.05）.However,the postoperative monitoring results showed that the postoperative stress level in the group with complications was significantly higher than that in the group without complications,and the differences were statistically significant（all P<0.05）.4.There were no significant differences in the intra-operative fluid intake and bleeding volume between patients with complications and without complications（P=0.303,P=0.271）.However,the daily fluid deficit in the patients with complications was greater than that in the patients without complications from the second day after surgery（all P<0.05）.Moreover,this study confirmed that patients with complications had a lower and statistically significant probability of fluid intake/output ratio<1 within4 days after surgery compared with those without complications[9（10.5%）vs91（36.5%）,P<0.001],and the final duration of fluid intake/output ratio>1 was longer（6.12 vs 3.86 days,P<0.001）.5.After univariate and multivariate Logistics regression analysis,independent risk factors related to the complications within 7 days after surgery were identified,including BMI≤25 kg/m²（OR=3.34,P=0.008）,T（temperature）≥37℃on the fourth postoperative day（OR=4.37,P=0.029）,N（neutrophil percentage）≥75%on the fourth postoperative day（OR=4.36,P=0.001）,FBG（fasting blood glucose）≥6.2 mmol/L on the fourth postoperative day（OR=10.11,P<0.001）,and the presence of fluid intake/output ratio<1 within 4 days after surgery（OR=5.38,P=0.004）;And independent risk factors related to the complications within 30 days after surgery,including BMI≤25 kg/m²（OR=5.52,P=0.002）,T≥37℃on the seventh postoperative day（OR=6.68,=0.048）,N≥75%on the seventh postoperative day（OR=5.24,P=0.005）,FBG≥6.2 mmol/L on the fourth postoperative day（OR=7.72,P=0.001）,FBG≥6.2 mmol/L on the seventh postoperative day（OR=3.45,P=0.027）,and（the final duration of fluid intake/output ratio>1）≥6 days（OR=14.00,P<0.001）.6.Based on Logistics regression analysis,nomogram model Nomogram-A was constructed using the above five variables to predict the incidence of complications within seven days after surgery.The AUC of Nomogram-A in the training set was 0.870（95%CI:0.822–0.917）,the sensitivity was 0.814,and the specificity was 0.813.The AUC in the validation set was 0.779（95%CI:0.638–0.920）,the sensitivity 0.820,and the specificity 0.706.Similarly,nomogram model Nomogram-B was constructed using the above six variables to predict the occurrence of complications within 30 days after surgery.The AUC of Nomogram-B in the training set was 0.930（95%CI:0.899–0.961）,the sensitivity was 0.845,and the specificity was 0.914.The AUC in the validation set was 0.892（95%CI:0.807–0.977）,the sensitivity was 0.826,and the specificity was0.857.7.Based on RF,SVM,and ANN,and by comparing AUC values from the algorithmic model validation set,we identified the random forest as the optimal model for predicting postoperative complications within 7 days after surgery(AUC_RF=0.754,AUC_SVM=0.737 and AUC_ANN=0.688).Further,using the random forest,we identified the top ten high-weighted variables that had the strongest correlation with the complications within 7 days after surgery,in order:FBG on the fourth postoperative day,N on the fourth postoperative day,BMI,T on the fourth postoperative day,the presence of fluid intake/output ratio<1 within 4 days after surgery,age,WBC（white blood cell）on the first postoperative day,WBC on the fourth postoperative day,cardiovascular disease,and T on the first postoperative day.Similarly,the random forest was the best algorithmic model for predicting complications within 30 days after surgery(AUC_RF=0.776,AUC_SVM=0.768,and AUC_ANN=0.737).The top ten high-weight variables were as follows:the final duration of postoperative intake/output ratio>1,FBG on the fourth postoperative day,FBG on the seventh postoperative day,N on the seventh postoperative day,N on the fourth postoperative day,BMI,T on the seventh postoperative day,age,cardiovascular disease,and FBG on the first postoperative day.8.Based on the algorithm of artificial intelligence,Nomogram-A’was constructed to predict the occurrence of complications within 7 days after surgery using the first 8high-weight variables related to complications within 7 days after surgery.The AUC value of Nomogram-A’in the training set was 0.888（95%CI:0.844–0.933）,the sensitivity 0.879,and the specificity 0.797.The AUC in the validation set was 0.805（95%CI:0.676–0.934）,the sensitivity was 0.860,and the specificity was 0.821.Similarly,Nomogram-B’was constructed to predict the occurrence of complications within 30 days after surgery using the first 8 highly weighted variables associated with complications within 30 days after surgery.The AUC value of Nomogram-B’in the training set was 0.937（95%CI:0.909–0.966）,the sensitivity was 0.909,and the specificity was 0.840.The AUC in the validation set was 0.895（95%CI:0.816–0.975）,the sensitivity was 0.857,and the specificity was 0.925.Comparing Nomogram-A and Nomogram-A’,the latter showed better AUC value in the ROC curves of training set and verification set(training set:AUC_Logistics 0.870 vs AUC_AI 0.888;verification set:AUC_Logistics 0.779 vs AUC_AI0.805).Similarly,comparing Nomogram-B and Nomogram-B’,the latter showed better AUC values in the ROC curves of the training set and the verification set(training set:AUC_Logistics 0.930 vs AUC_AI 0.937;verification set:AUC_Logistics 0.892 vs AUC_AI 0.895).It was proved that Nomogram-A’and Nomogram-B’based on artificial intelligence were more accurate in predicting postoperative complications.The calibration curves of Nomogram-A’and Nomogram-B’in the training set and verification set were constructed,and the actual prediction curve was close to the perfect curve,which showed that the model had good efficiency in the training set and verification set.Conclusions1.Postoperative complications are often accompanied by the increase of stress level,and the change of stress level is prior to the occurrence of complications.Early postoperative high stress state can predict the possibility of complications.2.Based on pathophysiology and stress generation mechanism,fluid intake and output can be used to reflect stress levels and serve as a potential factor for the occurrence of postoperative complications,providing a new idea for subsequent research on prediction of postoperative complications.3.Compared with the traditional Logistics regression analysis,the related factors of complications selected by artificial intelligence and the nomogram prediction model constructed have better prediction accuracy and good efficiency.