The use of artificial neural networks in decision support in cancer: A systematic review

Abstract

Artificial neural networks have featured in a wide range of medical journals, often with promising results. This paper reports on a systematic review that was conducted to assess the benefit of artificial neural networks (ANNs) as decision making tools in the field of cancer. The number of clinical trials (CTs) and randomised controlled trials (RCTs) involving the use of ANNs in diagnosis and prognosis increased from 1 to 38 in the last decade. However, out of 396 studies involving the use of ANNs in cancer, only 27 were either CTs or RCTs. Out of these trials, 21 showed an increase in benefit to healthcare provision and 6 did not. None of these studies however showed a decrease in benefit. This paper reviews the clinical fields where neural network methods figure most prominently, the main algorithms featured, methodologies for model selection and the need for rigorous evaluation of results.

Introduction

In the last decade, the use of artificial intelligence (AI) has become widely accepted in medical applications. This is manifested by an increasing number of medical devices currently available on the market with embedded AI algorithms, together with an accelerating pace of publication in medical journals, with over 500 academic publications each year featuring Artificial Neural Networks (ANNs) (Gant, Rodway, & Wyatt 2001). Claimed advantages of neural network methods include:

• Ease of optimisation, resulting in cost-effective and flexible non-linear modelling of large data sets.

• Accuracy for predictive inference, with potential to support clinical decision making.

• These models can make knowledge dissemination easier by providing explanation, for instance, using rule extraction or sensitivity analysis (Lisboa, 2002).

The published literature suggests that ANN models have been shown to be valuable tools in reducing the workload on the clinicians by detecting artefact and providing decision support, potentially with the ability to automatically re-estimate the model on-line. However, there are relatively few published clinical trials, and even fewer testing the clinical value of ANNs against established linear-in-the-parameters statistical methods (Lisboa, 2002).

There are two recurring concerns on ANNs. The first is the use of first principle statistical methods to control model complexity, which has been addressed by regularisation methods and with the use of cross-validation (Biganzoli et al., 1998, Lisboa et al., 2003, Ripley, 1996, Ripley and Ripley, 2001). The second key issue is transparency, i.e. explaining what influences the network predictions and how to resolve outcome predictions in terms of readily understood clinical statements. This is partly addressed by rule-extraction algorithms.

Notwithstanding these concerns, an interesting feature of neural network decision support in medicine is the routine clinical use of a range of systems, from the commercial-C.Net (Nabney, Evans, Tenner, & Gamlyn, 2001) and BioSleep (Tarassenko, McGrogan, & Braithwaite, 2002)—to research prototypes (Lisboa et al., 2000, Taktak et al., 2004) without listing in PubMed of supportive clinical trials. The situation is not specific to neural networks, but extends particularly to web-based decision support tools such as www.adjuvantonline.com, marking a departure from algorithms for clinical routine assessments, e.g. the Glasgow Coma Score for severity of illness in critical care and Nottingham Prognostic Index for breast cancer, both of which have undergone rigorous multi-centre clinical trials evidenced in the literature, if not altogether without controversy.

The use of unstructured approaches to clinical evaluation of new medical research is a trend, which has proved hard to change. Already in 1994 a paper entitled ‘the scandal of poor medical research’ (Altman, 1994) highlighted the need to proper study design bordering on the unethical typically through the application of such bad scientific methodology as to be sometimes called ‘torturing the data’ until they confess to the desired result (Mills, 1993).

Therefore, it is important to define and keep to a staged framework to design a sequence of studies each with a clear-cut purpose, ranging from the exploratory to the definitive, where the chief aim of each step in this chain is to support the next developmental step until a power calculation is possible which will determine the sample size, along with clinical protocol and study design for a multi-centre randomised clinical trial. Such a framework has been published (Campbell et al., 2000) and adapted for the development of intelligent decision support in an earlier review (Lisboa, 2002). This review will note the current trends in the studies that reach journals in the medical or medically related science literature, highlight points of good and poor practice, and draw conclusions for study design to improve the likelihood of studies being appropriately followed-up in the future.