Evolutionary assembled neural networks for making medical decisions with minimal regret: Application for predicting advanced bladder cancer outcome

• Published in: Expert systems with applications Vol. 41; no. 18; pp. 8092 - 8100
• Main Authors: Vukicevic, Arso M., Jovicic, Gordana R., Stojadinovic, Miroslav M., Prelevic, Rade I., Filipovic, Nenad D.
• Format: Journal Article
• Language: English
• Published: Amsterdam Elsevier Ltd 15.12.2014; Elsevier
• Subjects: Animal tumors. Experimental tumors; Applied sciences; Artificial intelligence; Artificial Neural Network; Biological and medical sciences; Bladder cancer; Calibration; Computer science; control theory; systems; Computerized, statistical medical data processing and models in biomedicine; Connectionism. Neural networks; Decisions; Evolutionary; Exact sciences and technology; Experimental renal and urinary tract tumors; Expert systems; Genetic Algorithms; Information systems. Data bases; Learning theory; Mathematical models; Medical management aid. Diagnosis aid; Medical sciences; Memory organisation. Data processing; Neural networks; Optimization; Regret Theory; Software; Tumors; Artificial Neural Network; Bladder cancer; Expert systems; Genetic Algorithms; Regret Theory; Measurement; Input output; Evolutionary algorithm; Epidemiology; Modeling; Goal programming; Urinary bladder; Cohort study; Medical application; High performance; Probabilistic approach; Decision support system; Health staff; Decision making; Calibration; Neural network; Decision analysis; Experimental study; Forecasting; Standards; Discrimination; Genetic algorithm; Follow up study; Activation function
• ISSN: 0957-4174; 1873-6793
• DOI: 10.1016/j.eswa.2014.07.006

•A novel two-step procedure for obtaining reliable ANN predictive models is presented.•Optimal configuration of ANN was performed automatically using Genetic Algorithms.•Clinical utility was estimated by integrating the Regret Theory Decision Curve Analysis into the procedure.•For predicting of advanced bladder cancer outcome soft-max activation functions and good calibration are the most important.•Compared to the alternatives better prognostic performances were achieved while user-dependency was significantly reduced. Development of reliable medical decision support systems has been the subject of many studies among which Artificial Neural Networks (ANNs) gained increasing popularity and gave promising results. However, wider application of ANNs in clinical practice remains limited due to the lack of a standard and intuitive procedure for their configuration and evaluation which is traditionally a slow process depending on human experts. The principal contribution of this study is a novel procedure for obtaining ANN predictive models with high performances. In order to reach those considerations with minimal user effort, optimal configuration of ANN was performed automatically by Genetic Algorithms (GA). The only two user dependent tasks were selecting data (input and output variables) and evaluation of ANN threshold probability with respect to the Regret Theory (RT). The goal of the GA optimization was reaching the best prognostic performances relevant for clinicians: correctness, discrimination and calibration. After optimally configuring ANNs with respect to these criteria, the clinical usefulness was evaluated by the RT Decision Curve Analysis. The method is initially proposed for the prediction of advanced bladder cancer (BC) in patients undergoing radical cystectomy, due to the fact that it is clinically relevant problem with profound influence on health care. Testing on the data of the ten years cohort study, which included 183 evaluable patients, showed that soft max activation functions and good calibration were the most important for obtaining reliable BC predictive models for the given dataset. Extensive analysis and comparison with the solutions commonly used in literature showed that better prognostic performances were achieved while user-dependency was significantly reduced. It is concluded that presented procedure represents a suitable, robust and user-friendly framework with potential to have wide applications and influence in further development of health care decision support systems.