TED Conversations

Marieke Jackson

This conversation is closed. Start a new conversation
or join one »

Neural Network for Behavioral Infectious Disease Risk Behaviors

I am a public health scientist - not a computer programmer or mathematician. I want to know what I need to study to feel comfortable discussing the design of a neural network that would be used to weight risk behaviors associated with transmission of STIs and HIV.

0
Share:
progress indicator
  • thumb
    Jun 13 2012: I too am a public health scientist and have worked with neural networks on public health questions, although not enough to make me an expert. Nevertheless, some advice might be helpful. For example, I used a neural network application in an HIV study:
    SAMPLE: Jordan, T. J. & Montgomery, R. L. (2000). Use of traditional analytic methods versus neural networks to assess risk: The example of HIV seropositive patients with pulmonary complications. Mathematics and engineering techniques in medicine and biological sciences, pp. 155-163. Las Vegas: CREA Press.
    The neural-net model did predict risk somewhat better than the traditional analytic approach, however this result did not seem to be any major improvement in predicting clinical effectiveness. The time and effort to create the model and prepare it for predictive analysis at the time the data were analyzed, over 10 years ago, was extensive. I am certain that new predictive modeling applications developed by SPSS and SAS expedite the creation of the model for the analysis, however the ‘tweaking’ of the model to ‘train it’ for maximum improvement can take a great deal of time. Further, neural-net models do not usually provide insight into the calculations that produced the result – they are a ‘black box’ that runs for hours or days, spit out a result, and shuts up. Both SPSS and SAS offer extensive manuals for their applications that can be very helpful, however hands-on experience and sweat equity are always important to get any outcome from a computer. Good texts are available online, also.
    Not knowing the type of data you have or its quality, or the model you have in mind makes any more specific response difficult. If you wish to share more information with me, I will try and help. In general, I do not believe that neural network models are for the novice or when under time pressure to get a result.