Free Trial

Safari Books Online is a digital library providing on-demand subscription access to thousands of learning resources.

  • Create BookmarkCreate Bookmark
  • Create Note or TagCreate Note or Tag
  • PrintPrint
Share this Page URL
Help

Chapter V. Data Mining Medical Informati... > Discussion and Conclusion

Discussion and Conclusion

Implications for Practice

Interpreting Table 5 is difficult. While three of the four ANN models show Abdomen AIS to be the most important factor in determining whether a patient lives or dies, it should be remembered that it is a network of factors that calculates the output, and Abdomen AIS can be used only to make a prediction in conjunction with all the other factors. As the following demonstrates, the ANN is not designed to analyse individual factors in the way that logistic regression can. Apart from ANN2, the models show reasonable agreement in the most important factors. ANN2, however, suggests that time, day, and month of admission are important determinants of death. This appears counter-intuitive, although when these factors are removed as inputs and then the ANNs are retrained, the predictive accuracy of each drops to around 73%. Clearly, therefore, these factors do have an impact on building the classifier. It is possible to suggest ideas about why this might be (staff will be affected by tiredness at difference times of the day, staff and patients may be affected by the seasons, and staff turnover may have an impact from year to year). It seems far-fetched to suggest that these have a real impact on mortality, and in fact, when chi-squared tests were performed to examine the relationship between these factors and death, no significant association was found. So, it is known that these factors have some influence, but it is not clear what that influence is. Perhaps there are complex cross-correlations between these factors and others that can be modelled by the ANN but not explained. Discussions on how to deal with peculiarities like this are scarce in the ANN literature. One way of trying to examine these cross-correlations is to list the meaningful factors with which day, month, and year could be correlated. In this instance, there do not appear to be any. This goes back to one of the two big disadvantages of using ANNs: it is not possible to show why a neuron is weighted in a certain way after training. Without this, the usefulness of ANNs is severely impaired, and, in the case of trauma audit data, illustrates the limitations of an ANN as an analysis tool.

While ANNs may not be suitable alone for analysing trauma audit data, they may have a place alongside more traditional techniques. The results demonstrate the value of recording as much data about the trauma as possible and including it in an exploratory analysis. An analysis by ANN may be valuable for identifying factors that previously have not or normally would not be identified as having an impact on outcomes, such as time of day, which was identified in this study by an ANN but not by logistic regression. These factors then can be investigated by traditional techniques in order to examine their impact.


  

You are currently reading a PREVIEW of this book.

                                                                                                                    

Get instant access to over $1 million worth of books and videos.

  

Start a Free 10-Day Trial


  
  • Safari Books Online
  • Create BookmarkCreate Bookmark
  • Create Note or TagCreate Note or Tag
  • PrintPrint