Friday, June 24, 2005

Utility of neural network in assessing colon cancer prognoses.

Networking - Combating cancer

714 words
25 June 2005
The Economist
(c) The Economist Newspaper Limited, London 2005. All rights reserved

Combating cancer with neural networks

Picking the best treatments for cancer patients

WHILE particle physics can be esoteric, its practitioners are keen to show it has practical applications. They invented the world wide web. They also contributed to a number of advances in medicine, among them positron-emission tomography, a body-scanning technique. Now, as if to strengthen the case, a group of particle physicists led by Robin Marshall of the University of Manchester, in Britain, has applied its knowledge of information technology to show how computer programs known as neural networks can help doctors to choose the best treatments for people with cancer.

Unlike a conventional computer, which takes data, processes it using an algorithm and generates a definite answer, a neural network learns to create a range of answers from a range of inputs. To do this, it is “taught” by being fed a series of training inputs and then told what the answer should be in each case. The network adjusts the weighting of its internal connections to try to retain the correct matches as far as possible. Once the teaching process is complete, the network can be used to calculate answers from new inputs.

Like many in his field, Dr Marshall uses neural networks to discard the huge amounts of boring data produced in particle colliders and to identify the interesting events. The network learns to associate a particular range of inputs with interesting collisions and to ditch the rest. He has now turned this expertise to the medical field, following a chance meeting with Sir Alfred Cuschieri, an oncologist at Ninewells Hospital in Dundee.

Ninewells has detailed records on thousands of patients with colorectal cancer. These records contain a wide range of information such as each patient's age, sex, type of treatment, size of tumour and eventual fate. Dr Marshall realised that a neural network could be trained with this information to calculate the survival chances of other people with the same condition. The machine would learn to associate certain ranges of patient profiles with particular survival probabilities.

Dr Marshall and Sir Alfred, together with some colleagues from Manchester and Dundee universities, selected those records that contained enough data to create a detailed profile of a patient at the beginning of his treatment, and to follow his progress over the subsequent five years. They then used 1,558 of these records to train a neural network. In each case, the input was 16 pieces of data that defined the state of the patient. The output (ie, what the network was trying to learn to predict) was the patient's fate—in other words, whether he died over the course of the five years and, if so, when.

The researchers, who will publish their work in a forthcoming issue of Concurrency and Computation: Practice and Experience, then tested the trained network by using the same 16 parameters from each of the remaining 1,220 records as the input, while withholding information about the survival of the patient. They found that, in 90% of cases, the time at which the neural network predicted that a patient's chances of survival would fall below 40% was within three months of the actual time of death of that patient.

According to Sir Alfred, this system is ideally suited to predicting the survival chances of individuals. He says that the statistical techniques currently used by doctors to calculate a person's chances of surviving a disease such as cancer are a blunt instrument. By contrast, the neural network created at Manchester enables them to give individual prognoses, so they do not have to rely on crudely defined average chances of survival.

Once the project's researchers have verified the reliability of their neural network, they intend to make it accessible over the internet. Doctors will be able to enter their patients' parameters and generate prognoses. More importantly, they will be able to compare the effects of different treatments by varying the relevant inputs.

The researchers also believe that their system could be applied to the treatment of a variety of other chronic disorders, such as heart disease and diabetes. That would create further evidence that particle physicists do live in the real world, at least some of the time.


Post a Comment

<< Home