Prediction of postoperative morbidity after lung resection using an artificial neural network ensemble
Introduction
Lung resection is currently the best therapeutic option for patients with localised bronchogenic carcinoma. In the last years, a number of articles have been published in which several independent variables frequently associated to lung cancer, such are chronic obstructive pulmonary disease (COPD), cardiac ischemia, and others, are correlated to mortality or morbidity after major pulmonary resection [2], [3], [12]. Although knowledge of these variables can help to prevent postoperative morbidity, the applicability of these data to individual decisions has not been demonstrated, since constructed predictive regression models usually do not have a high sensitivity or specificity [12], [23].
In order to increase the performance of risk-prediction models in thoracic surgery, some investigators have tested single neural networks. In the reports by Lippman and Shahian [19] and Tu et al. [22], the measured outcome was in-hospital mortality after coronary artery by-pass graft. The capacity of logistic regression (LR) and artificial neural network (ANN) models to predict the outcome was similar in both publications. To our knowledge, there is only one published report in which single neural networks has been applied successfully to predict the risk of lung resection in a small series of cases [8].
Combining the prediction of multiple classifiers to produced a single classifier (known as an ensemble) has been demonstrated to increase the accuracy and performance of prediction [20], [24], [25]. In this report we evaluate the performance of an artificial neural network ensemble to predict cardio-respiratory morbidity after pulmonary resection for non-small cell lung cancer (NSCLC).
Section snippets
Study population
From January 1994 to December 2001, 515 patients underwent lobectomy or pneumonectomy for NSCLC in our hospital. All cases have been operated upon by the same surgical team. Selection criteria for operation consisted in the absence of major co-morbidity refractory to medical therapy, pO2 at rest over 50 mmHg, pCO2 under 46 mmHg and postoperative FEV1% (ppoFEV1%) over 30% of the normal value. Calculation of the ppoFEV1% was based on the number of non-obstructed pulmonary segments to be resected
Results
Table 3 depicts the results of descriptive analysis of continuous variables in both subsets of patients. Significant differences were found for age (series B patients were slightly older), preoperative chemotherapy (more cases in series B underwent induction treatment) and pathological tumour staging (more advanced tumours in series A).
Overall mortality in the two populations was 29 cases (5.9%): 24 cases in series A (6.9%) and 5 in series B (3.5%;P=0.205). Overall cardio-respiratory morbidity
Discussion
Pulmonary resection remains the most effective therapy in patients having localised lung cancer. Due to its relation with tobacco abuse, lung cancer patients frequently have other associated cardio-pulmonary diseases such are COPD or cardiac ischemia, which are related to postoperative morbidity and mortality.
In the last years, several single-institution [2], [3] or multi-institutional reports [7], [12], [16] have been carried out to identify clinical predictors of morbidity or mortality after
Conclusion
In conclusion, in our series, artificial neural network ensemble offers a high performance for the prediction of cardio-respiratory morbidity. Our results must be considered with caution and their future applications for decision making in individual cases of lung resection surgery remain to be investigated.
Acknowledgements
We are indebted to anonymous reviewers who have greatly improved the manuscript with their suggestions.
References (25)
- et al.
Identification of prognostic factors determining risk groups for lung resection
Ann Thorac Surg
(2000) - et al.
Pneumonectomy for malignant disease: factors affecting early morbidity and mortality
J Thorac Cardiovasc Surg
(2001) - et al.
POSSUM Scoring System as an instrument of audit in lung resection surgery
Ann Thorac Surg
(1999) - et al.
Early complications in surgical treatment of lung cancer: a prospective, multicenter study
Ann Thorac Surg
(1997) - et al.
Neural networks as a prognostic tool of surgical risk in lung resection
Ann Thorac Surg
(2002) - et al.
A decade of change-risk profiles and outcomes for isolated coronary artery bypass grafting procedures, 1990–1999: a report from the STS National Database Committee and the Duke Clinical Research Institute, The Society of Thoracic Surgeons
Ann Thorac Surg
(2002) - et al.
Receiver operating characteristic curve analysis of clinical risk models
Ann Thorac Surg
(2001) - et al.
Prospective analysis of pneumonectomy: risk factors for major morbidity and cardiac dysrhythmias
Ann Thorac Surg
(1996) - et al.
Prognostic models of thirty-day mortality and morbidity after major pulmonary resection
J Thorac Cardiovasc Surg
(1999) - et al.
Assessment of operative risk in patients undergoing lung resection
Chest
(1994)
Coronary artery bypass risk prediction using neural networks
Ann Thorac Surg
Lung cancer cell identification based on artificial neural network ensembles
Artif Intell Med
Cited by (46)
Integration of artificial intelligence in lung cancer: Rise of the machine
2023, Cell Reports MedicineBoosted SVM for extracting rules from imbalanced data in application to prediction of the post-operative life expectancy in the lung cancer patients
2014, Applied Soft Computing JournalCitation Excerpt :Taking into account the limitations of both classic statistical approaches and hospital datasets, which are often incomplete, due to missing values of attributes and unknown survivals for prospective analysis, and also – aiming at increasing not only formal prediction accuracy (which can easily be high with heavily imbalanced data) but also other performance measures like AUC or Gmean, other multivariable approaches have been proposed for the TS domain. Those include data mining and machine learning procedures for standardizing TS data analysis and report generation [51,52], as well as applications based on particular techniques, like decision trees [53,54] or artificial neural networks [55,56]. In this chapter we report application of the proposed approach in modelling surgical risk for real-life clinical data from Thoracic Surgery domain.
Can we predict morbidity and mortality before an operation?
2013, Thoracic Surgery ClinicsCitation Excerpt :These investigators developed models for estimating mortality and morbidity that had 100% specificity and sensitivity in classifying 28 cases. Santos-García and colleagues64 examined neural networks to predict cardiopulmonary morbidity after pulmonary resection for NSCLC by using a training set of 348 patients and then validating this in 141 cases. Their neural network ensemble model combined 100 separate neural networks with different structures and parameters and showed a remarkably high predictive performance, with an ROC c-statistic of 0.98.