Description
Objectives: The aim of this study was to compare different classification methods for the retrospective detection of patients with one or more surgical and/or medical complications. The secondary aim was to show the potential of a structured search of hospital data for the classification of special patient groups. Methods: Data from hospital data bases of eight clinics in Upper Austria were combined with data collected by data nurses from handwritten patient documentation. Complication (yes/no) was selected as the target variable. Patients undergoing acute appendectomy (n= 1341), acute and elective cholecystectomy (n= 1307) and elective inguinal hernia repair (n = 1574) were included into the study. Data on laboratory as well as radiologic, serologic, histologic and microbiologic services received by the patients during and 42 days before and after their stay was combined with 45 parameters concerning patient risk. The data defining patient risk included operation type, ASA-score, BMI, sex, age, renal-, pulmonary-, cardiologic-, gastrointestinal- and “special-” risk factors defined individually for each operation. For each of the three datasets three different methods, logistic regression, boosting for generalized linear models (boosting) and classification trees (ctree) were applied to detect patients with complications. The models were fitted using the statistical analysis software R. Data of patients undergoing appendectomy (n= 196), cholecystectomy (n= 197) and inguinal hernia repair (n= 210) were used as a test data set to evaluate the finally trained models of each operation. To assess the predictive performance of the models, positive predictive value (PPR), rate of positive predictions (RPP) and sensitivity were compared. In addition to predictive performance the number variables in the final model is also an important criterion for choice of the final model as more variables need more efforts on data collection and preparation. Results: A comparison of sensitivity, PPV and RPP showed that models selected by boosting and ctree perform better than the classical logistic regression models. For inguinal hernia repair data, the boosting (RPP = 36.7 %, PPV = 14.3 %, Sensitivity = 91.7 %) and the ctree model (RPP = 32.4 %, PPV = 16.2 %, Sensitivity = 91.7 %) lead to similar results with cut-off levels at 0.03 for the boosting model and 0.07 for the ctree. For cholecystectomy, the ctree (RPP = 28.9 %, PPV = 31.6 %, Sensitivity = 94.7 %) outperformed the boosting model (RPP = 22.8 %, PPV = 25.6 %, Sensitivity = 84.2 %) in terms of PPV and Sensitivity. The cut-offs were at 0.08 and 0.10. For appendectomy data, the ctree (RPP = 29.1 %, PPV = 22.8 %, Sensitivity = 74.7 %) outperformed the boosting model (RPP = 37.8 %, PPV = 18.9 %, Sensitivity = 82.4 %) in terms of RPP and PPV. For all three operations the final models using ctree are rather small, involving only 9, 6 and 7 variables for hernia, cholecystectomy and appendectomy respectively compared to 10, 16, and 12 in the boosting model. Conclusion: Application of our classification tree models on existing data can save time and costs for the identification of patients with a history of surgical and medical complications in surgery.| Period | 3 Oct 2017 |
|---|---|
| Event title | Isqua's 34th international Conference |
| Event type | Workshop |
| Location | London, AustriaShow on map |