Digital Health für Einsteiger

1. 00:00Welcome back to "Digital Health for Beginners", the block on AI.
2. 00:06So far, we've been dealing primarily with Machine Learning and Deep Learning as part of Machine Learning.
3. 00:12Today I would like to say a little bit more about classical AI methods, namely expert systems.
4. 00:19The bottom line is that expert systems aim to, answer complex questions and make important decisions.
5. 00:31And they do that by means of if/then rules, so by means of logic.
6. 00:37That is, they are based on logic and they use quasi queries on, for example, in the case risk contacts, fever, cough,
7. 00:56to then come to a decision, for example, whether or not to order a lab test for a covid patient.
8. 01:05And they end up doing it by asking a lot of these if/then questions, that are put together by experts, that's why they're called expert systems.
9. 01:18So for example, did someone have a risk contact, yes, he also has a fever, yes, then you should definitely order a lab test.
10. 01:29But he doesn't have a fever, then maybe it's not so bad after all, but he also has a cough, which again is a different system,
11. 01:44then the risk goes up again, and that's why you say, okay, then one would do it after all.
12. 01:48These expert systems are actually everywhere, and simple expert systems really integrated everywhere,
13. 01:57without really calling them AI specifically now, because in the end they are just if/then queries.
14. 02:03Only, these very simple, trivial, if/then queries can be just extremely complex, when you connect several together.
15. 02:12So it can get extremely complex very quickly and really for very complex queries, they can also be somewhat inflexible.
16. 02:22By the fact that they are defined by experts, it is of course also a very high workload for experts.
17. 02:30An alternative to expert systems, which are in the end a generalization, are so-called Bayesian networks.
18. 02:40Logic, on the one hand based on 1/0, so binary, but in the real world, there's just uncertainty.
19. 02:51So we rarely know about all things, do they apply or not, but we can give probabilities.
20. 03:00In logic it is like this, one would say from A follows B, it's raining, so the road is wet.
21. 03:13But that doesn't allow us now, for example, to say, okay, if the road is wet, what is the probability that it has rained, because the road could also be wet for other reasons.
22. 03:26However, if we think about it, a lot of times when the road is wet, it was raining before.
23. 03:34That's why when you work with probabilities, so you can't say, okay, we're using logic now to say it's raining, period, because the road is wet.
24. 03:43Can we say, okay, given that the road is wet, the probability of it raining is ten percent, for example.
25. 03:53So probability theory in general Is a generalization of logic.
26. 03:59And these uncertainties that you have, are represented as numbers between 0 and 1,
27. 04:04where 0 means in the end that one is quite sure for wrong, quasi the event does not apply and 1 means that one is quite sure that an event is true.
28. 04:21And so you can nest different events, so to speak and complex interactions similar to this logic-based expert system, that we just saw in Bayesian networks.
29. 04:39In which each node is just such a variable, such an event, and these nodes are connected by arrows, so they are connected in networks, so to speak.
30. 04:56And these arrows are directed, quasi A points to C, for example, Sugar consumption points to diabetes,
31. 05:06We would sort of draw up a network here like this, where we say, okay, the probability of diabetes is conditioned by two other components,
32. 05:19one is sugar consumption, and the other is for example genetic risk for diabetes, a genetic predisposition.
33. 05:28And if someone eats a lot of sugar, the likelihood of diabetes goes up,
34. 05:33if he has a genetic predisposition, this probability also goes up and we can show that in this graph.
35. 05:41The interesting thing now, as opposed to through this, that we now work with probabilities and we don't now say, okay, it all has to be 0 or 1,
36. 05:53but there can also be values in between, we can make such reversal conclusions.
37. 05:58So if we go now again for example, to it is raining, the road is wet and represent that as a Bayesian network, so we sort of have one node for "it's raining" and one for "the road is wet",
38. 06:14then we can quas on the one hand, if it is raining, we can say something about the probability of road is wet, that is the probability of B given A.
39. 06:28But we can also, again using Bayes formula. just as we do in the Naive Bayes classification procedure, to say something about a given b.
40. 06:47So that's sort of something that we can do with probabilities, in that we don't have to say if the road is wet, it's always rained,
41. 06:54but we can say, okay, then the probability is just ten percent that it has rained.
42. 06:57And the way we work that out is again with this formula, which is just the reversal probability, probability B given A,
43. 07:05so the probability that the road is wet, given that it rained times the probability that it rained at all divided by the probability that the road is wet.
44. 07:16And we can use these Bayes nets just as well again for our problem as before, by virtually connecting these variables into nets.
45. 07:35So our question is whether or not somebody has covid or maybe has a normal cold or not,
46. 07:48then we can, we can include those as variables in this network as well, in addition to the other things that we just queried.
47. 07:57and these individual variables just with conditional probabilities, which are represented by these arrows.
48. 08:08What we see is that this is some of these variables that are all in the same net, some of them are blue, some of them are gray.
49. 08:17The reason is that these blues are observed variables, the grays are so-called hidden, latent variables that are not observed, but which we would like to model anyway.
50. 08:32So we are interested in saying based on the observations, what is the probability of the unobserved variables, so for example, whether someone has a covid infection or has a cold.
51. 08:45And how do we do that? We just do that exactly by applying Bayes formula and use it to calculate these conditional probabilities.
52. 08:55That's how expert systems work, they allow us to sort of, answer questions and solve complex decision problems.
53. 09:09And they can also be generalized with probabilistic methods, and then they are called Bayes net.
54. 09:18And then they tell us something about the uncertainty of these individual binary variables.
55. 09:25With that, I would like to end already this short entry into classical AI again,
56. 09:35before we in our last part on AI we would like to briefly discuss risk factors and ethics.
57. 09:46See you soon.

To enable the transcript, please select a language in the video player settings menu.