AI and decision making

Rehabilitation decisions are complex because so many factors have an influence and we do not know the extent of the influence of each factors, nor do we know about interactions between factors, and we cannot easily compute the final outcome. Artificial intelligence (AI) might help. An interesting paper, just published, shows in a well-designed randomised trial that using AI to predict outcomes, coupled with information about patient preferences, can assist in making a decision, not only increasing satisfaction but leading to a better functional outcome.

The trial recruited 145 patients with osteoarthritis of the knee: 69 went through an educational programme and then saw the surgeon to discuss having a total knee replacement; 76 went through a ‘preferences module‘, determining what the patient considered important in terms of outcome and their opinion on risks associated with surgery. The data were then used by the AI-assisted computer programme to provide information to be used when they saw the surgeon.

Unfortunately the description of exactly what the AI did, and what data were given to the computer are not described in detail. My understanding is that the AI was trained on a large data-base to predict outcome after surgery or not having surgery. Therefore it was able to predict likely outcome (such as pain, mobility, complications) for the individual patient. I think that the programme also had access to patient preference and attitude data.

In the shared decision-making process, the surgeon used the outcome of the AI and preference data to discuss the patient’s choice. The discussion was not longer.

A second interesting (to me) aspect of this trial was the use of a simple measure of satisfaction with the shared decision making process. This was done using the CollaboRATE survey.

The paper describing the development of this three-question measure is interesting. I will quote its key finding, because I think it is of great importance to the development of all measures.

The key finding of this study is the confirmation that the correct end-user interpretation of a brief patient-reported measure of shared decision making is significantly improved by including the views of lay people in the development process, leading to the avoidance of terms such as ‘decisions’ and ‘preferences’.

Developing CollaboRATE: A fast and frugal patient-reported measure of shared decision making in clinical encounters. Elwyn et al, 2013

Artificial intelligence, if I understand it correctly, is fed data about factors that may be related to the outcome, and is fed data about the outcome. A neural network then uses the input data about outcome to learn what items are associated with what outcomes. It can automatically learn about interactions between two or more input factors.

So our responsibility, in rehabilitation, is to ensure that:

  • all potentially important input factors are considered and given to the artificial intelligence, and
    • the best possible measures are used for each factor chosen
  • all potentially relevant outcomes are considered, and the artificial intelligence is trained for each, and
    • the best possible measures are used for each outcome selected
  • all likely preferences and risk factors are considered when an educational ‘preferences module’ is drawn up.

The biopsychosocial model of illness will be extremely useful when considering the data needed, and also in structuring preferences.

My conclusion is that artificial intelligence can be used to develop much better predictions of outcomes that humans can, but humans need to ensure that the neural network is given appropriate data both in terms of relevant determining factors and in terms of relevant outcomes. Rehabilitation experts will be best placed to determine what data should be given to artificial intelligence programmes.

Make your comment or suggestion here

Scroll to Top
%d bloggers like this: