This repository contains all the code and results used in this study.
Postoperative complications following cancer surgeries are still hard to predict despite the historical efforts towards the creation of standard clinical risk scores. The differences among score calculators, contribute for the creation of highly specialized tools, with poor reusability in foreign contexts, resulting in larger prediction errors in hospital practice. This work aims to study and predict postoperative complications risk for cancerpatients, offering two major contributions. First, to develop and evaluate amachine learning-based risk score, specific for the Portuguese population using a retrospective cohort of 847 cancer patients undergoing surgery, predicting 4 outcomes of interest: i) existence of postoperative complications, ii) severity level of complications, iii) postoperative mortality within 1 year, and iv) number of days in the Intermediate Care Unit (ICU). Second, to support the study with relevant findings and improve the interpretability of predictive models. In order to achieve these objectives, a robust methodology for the learning of risk predictors is proposed, offering new perspectives and insights into the clinical decision process. For this cohort, the postoperative complications can be predicted with 0.69 AUC, complications’ severity with a 0.65 AUC, the days inthe ICU with a Mean Absolute Error of 1.07 days, and one-year death with 73% AUC. An external validation with independent data was carried, suggesting the models have the potential to guide and identify cancer patients at increased surgical risk.
The content found in this repository corresponds to all the initial data processing, which is then fed to a wide set of ML models. These models were improved through a development process, including resampling, normalization, hyperparameter optimization, feature selection, etc. The results were subject to succesive evaluation processes, in order to gather as much insight into models' performance as possible. Additionally, the final models were tested in an independent set of 137 patients, further guaranteeing their generalization ability.
Access to the original dataset is restricted due to data privacy concerns. To know more about the dataset used in this study, please contact:
- Daniel Gonçalves, dmateusgoncalves@tecnico.ulisboa.pt
- Rafael Costa, rafael.s.costa@tecnico.ulisboa.pt
- Rui Henriques, rmch@tecnico.ulisboa.pt
This repository contains the code files used in the article:” Gonçalves, D., Henriques, R., Santos, L. L., Costa, R.S. On the predictability of postoperative complications for cancer patients: a Portuguese cohort study. BMC Medical Informatics and Decision Making, 21, 200 (2021) | doi: 10.1186/s12911-021-01562-2.