We can reduce the number of clinical tests carried out by regulation, by rationing, or by analysis. Currently, we have very high cost analysers – doctors – who rely on lots of tests to make or confirm their diagnoses. If we produce a reliable analyser at low cost, we can save on doctor labour and we save on tests.
We could considerably increase the number of doctors and allow competition to reduce the cost of each doctor – however, the cost of training is significant, so costs would not halve, and the relentless increase of Medicare costs would continue (and if they are poorly trained, they will use even more tests).
Are we at the stage where we can create an automated system that learns from text, so the means of education of the system is not by programmers, but from the same text that is used to train doctors? Such an approach would have a dramatic effect on costs. It may be slower than a good doctor, but has the property of being easily copied, so every patient in a hospital can have their own machine accepting all relevant input regarding the particular patient (and effectively worrying about them).
How would such a system be any different to an expert system? The expert system runs on rules – IF A THEN B – which is a very poor (and inefficient) way to store knowledge. The proposed system would use an undirected and active structure. The active structure generates an algorithm within itself to solve any particular problem – the process may involve changing the structure that is generating the structure.
Ontologies have been tried – isn’t this just another one? An ontology is a poor way to store knowledge – it relies on rigid classifications, whereas in reality, many things are on the border between simple classifications, or move between them depending on association – it is necessary to use relations (the logic of relations) to describe the interconnectedness of things. More than that, it is necessary to use existential, sentential, temporal and relational logic all at once, which is something that people do poorly when it becomes complicated (so they make mistakes – "we are only human").
What would we need to demonstrate to show this approach could be successful?
·
Reading of text and creation of structure – particularly text that describes concepts at different levels. We can show this now in reading specifications.·
Integration and reuse of concepts.·
Sheer volume of knowledge required. We will assume specialisation, but would anticipate nodes in the tens of millions – this is possible, where we have a text reading facility and a medical knowledge (active) structure.
Related Material