UPM Institutional Repository

Single and multiple time–point artificial neural networks models for predicting the survival of gastric cancer patients


Dezfouli, Hamid Nilsaz (2016) Single and multiple time–point artificial neural networks models for predicting the survival of gastric cancer patients. Doctoral thesis, Universiti Putra Malaysia.


The extensive availability of recent computational models and data mining techniques for data analysis calls for researchers and practitioners in the medical field to opt for the most suitable strategies to confront clinical prediction problems. In many clinical research work, the main outcome under investigation is the time until an event occurs. Survival models are a collection of statistical procedures used to analyse data where the time until an event is of interest. Particularly the application of a data mining method known as ‘neural networks’ offers methodological and technical solutions to the problems of survival data analysis and prognostic model development. In this context, artificial neural networks (ANN) have some advantages over conventional statistical tools, especially in the presence of complex prognostic relationships. ANN model applications for modeling the survival of gastric cancer patients have been highlighted in a number of studies but without a full account of censored survival data. The primary task under investigation in this thesis is to develop neural network methodologies for modeling gastric cancer survivability and fill the gap in the current literature by adopting strategies that directly incorporate censored observations in the process of constructing a neural network model. The dataset used in the study comprises of patients with confirmed gastric cancer who underwent surgery at the Cancer Registry Center of Taleghani Hospital, Tehran, Iran. To achieve the research aims, single and multiple time-point ANN models are proposed. The first model is a single time-point ANN designed to predict the survival of patients at specific time points. The second is a multiple time-point model specifically designed to provide individualized survival predictions at different time points. Thus, an individual survival curve can be generated for a particular patient by plotting the survival probabilities produced by output units, which render the system more useful in clinical settings. The third model is a softmax ANN designed to estimate the unconditional probability of death and predict the time period during which death is likely to occur for an individual patient. All models are extended to incorporate censored data. Employing the strategies for imputing the eventual outcome for censored patients has allowed all the available data to be used in developing an ANN predictor model. Several criteria are employed to validate the models. The research demonstrated how ANNs can be used in the survival analysis for predictive purposes without imposing any restricting assumptions. The proposed models provide accurate predictions of survival with high levels of sensitivity and specificity. Additionally, the sensitivity analysis provided information about the relative importance of each input variable in predicting the outcome. To sum up, The ANN survival models presented in this thesis provide a framework for modelling survival data with censorship and facilitate individualized survival predictions. The findings will provide physicians and medical practitioners with information to improve gastric cancer prognosis and may assist in the selection of appropriate treatment plans for individual patients as well as efficient follow-up planning.

Download File

IPM 2016 15 IR.pdf

Download (1MB) | Preview

Additional Metadata

Item Type: Thesis (Doctoral)
Subject: Stomach - Cancer
Subject: Stomach - Diseases
Subject: Neural networks (Computer science)
Call Number: IPM 2016 15
Chairman Supervisor: Mohd Rizam Abu Bakar, PhD
Divisions: Institute for Mathematical Research
Depositing User: Haridan Mohd Jais
Date Deposited: 07 Mar 2019 07:31
Last Modified: 07 Mar 2019 07:31
URI: http://psasir.upm.edu.my/id/eprint/67515
Statistic Details: View Download Statistic

Actions (login required)

View Item View Item