Artificial Intelligence Detection of CPPD on Hand Radiographs Save

I'm Anthony Chen. I'm a consultant rheumatologist from Reading, United Kingdom, and I'm here in EULA twenty twenty four in Vienna. And here at the conference, there's been a lot of new developments in the field of digital rheumatology, particularly looking at areas such as machine learning and artificial intelligence, and how we can leverage that into the whole field of rheumatology, from the diagnosis to the management and to the prognostication for to help with our patient care. And there was a session that ran here at EULA twenty twenty four looking particular in this topic. And I'm very happy that we are joined today by Professor Thomas Hugo from the University Hospital Lausanne in Switzerland.

And he did a very nice oral presentation, the number is OP0112, where he used a model of calcium pyrophosphate deposition on hand radiographs and how deep learning model could be used for an automated detection of patients with CPPD. So Thomas, thank you for coming to join us here in RheumNow today. I wonder whether you could tell us, introduce us to the area of your presentation, and then maybe in the second half we can kind of look into the future as well as to how this whole field would go forward.

Thank you Anthony, it's a pleasure. So you know the idea of this study was based on the new classification criteria for CPPD from last year 2023, the EULA ACR criteria for the classification of chondrocalcinosis or calcium parasite deposition disease. So you can diagnose CPPD also without proof of crystals in the joint fluid so by clinical parameters and radiological findings for example or ultrasound findings. On And hand radiographs you can find quite a lot of signs obviously we know this from the clinic for CPPD like TFCC calcification but also MCP2 or three calcification or osteoarthritis specific sites. So we trained a model to recognize calcification from TFCC or MCP three or two on Hendrik radiographs and that's labeled by two independent radiologists by my colleague Fabio Bece from Lausanne.

And then we trained the model in almost 1,000 radiographs. Of those, around 300 positive for TFCC. We trained a deep learning classical convolutional neural network to recognize either TFCC calcification, MCP2three calcification, or a combined model with all three of them. And the findings were clearly that the combined model detecting TFCC but also MCP-two zero three classification is the best. We got an accuracy of 0.85 which corresponds to, let's say, a sensitivity of 77%, specificity of 80% in a relatively small data set.

We also then looked at it as specific sites of the lesions and did the five fold cross validation to see how it is. But the best is actually a combined model followed by a TFCC model for the detection of CPPD. We also showed some heat maps. Heat maps are very interesting, I think also for clinical implementation because they show you where, let's say, the region of interest for the algorithm, so the area on the radiograph where the algorithm took its decision, I think heat maps will come into clinical practice to help us, to guide us the way to diagnosis, if you wish.

So very promising results obviously, as you say, still early days and small number. What was the area under the curve when you did the the the calculation? Because increasingly, it's one of those areas where we kind of now have to look at sensitivity and specificity and also the AUC.

Yeah, it was 0.86 and we also calculated the area under the precision recall to be a bit more precise, because it was an imbalance data set, meaning more CPPD negative than positive, and so we used the AUPR which was 0.77 and again best for the combined model followed by the TFCC side and then much weaker versus MCP two and three. Mean that's what we also see in the clinic.

So for our listeners, these seem to be the new parameters now where we are increasingly being exposed to, this whole area of evaluating the, the sensitivity and also the specificity of this, particular test, and then looking at the, area under the curve. Now, very promising results. Where do you think along the the patient pathway will we be using this? In the early stage of diagnosis or when the diagnosis becomes more challenging? Because it is can be quite a challenging diagnosis to make CPPD sometimes we are confused with maybe say a different type of rheumatic condition.

That's a good question. The other rationale to do this study was actually to have a tool to screen other larger databases. Okay. So we did already did this. It's notably interesting in RA.

So we used it to screen our national registry data set with around 12,000 RA patients, but we cannot just score those images for CPPD. So the algorithm is very helpful and we found that almost twenty five percent of our RA patients were actually CPVD positive. So in first line, we want to use it to screen it in other databases or clinical trials, whether CPVD is confounder, for example, in RA. This is very interesting. And then in clinical practice, I think either way it's probably more a bit in the elderly patients where you know patients over 80 almost fifty percent has somehow CPPD.

I think it will be maybe more on the later stage. I think we will combine CPPD detection also with RA detection and OA detection or PSA detection and I think notably in combination with other models for RA, PSA etc this model will be interesting as a diagnostic tool potentially.

And how do you see this, particularly in that patient group where they're a bit older, maybe more females as well? How did the model work, you know, do you, you know, project that the model will be able to differentiate between OA which is probably very highly prevalent in that population with the CPPD?

That's a good question. Here this model I presented was purely on calcification, right? We have published another model to detect DRP osteoarthritis, so that's the next step actually. We want to combine the calcification model with an OA detection model, so we know exactly what is calcification, what is OA, and then potentially also with clinical data to give predictions.

So you're training a model but then the eventual application may be beyond CPPD. It's in other areas where where it can also be refined and Clinically sharpen up as well. Yeah. Yeah. So here at the EULA, there were a lot of other interesting presentations, certainly in your session, which we all enjoyed attending.

I wonder whether you wanted to share with us some of your highlights from maybe that session that you were presenting at.

Yeah, so a lot was imaging and I mean that reflects a bit the over 800 FDA approved AI algorithms. Most of them are in imaging and that was also the case in this session. There were different types, for example predicting progression of knee osteoarthritis by MRI or actually in arthralgia the use of MRI findings to predict the development of arthritis. That was interesting. One highlight was actually a technical issue that one speaker showed first relatively good results and then he pointed out a technical problem how to interpret the results of the algorithm and showed that it was actually not that good.

So they had to retrain it and so that gave me some methodological insight. There was one paper on large language model for patients receiving joint replacement, actually also a methodological paper to design questionnaires, then to assess patient satisfaction. I think those were the ones and another one on bone remodeling markers and then to predict actually inflammatory arthritis. I found that was also very interesting.

Yeah it seems to be in a quite a lot of areas now where certainly in my sort of and especially area of interest in spondyloarthropathies today there was a lot of good poster presentations but particularly looking to some of these areas as well so from early diagnosis to prognostication of patients response to the treatment by using, you know, machine learning algorithms to to work on to that. So any other things you'd like to kind of share with us from this session in the whole field of AI and machine learning before we sort of close?

Imaging is number one. I think simple tasks such as detecting calcification will be automized or you will have support by AI models. Clinical predictions are more difficult. We have seen this. I think that the AI models could help maybe in the shared decision process.

They're too complicated just to rely on, especially when it comes to treatment decision, which type of drug you should take. So I have nothing seen that tells me which type of drug we should use. So we're waiting for this kind of study, but it's promising. A lot will be integrated in the electronic medical health record. I think that's important to have those algorithms support us and save time, not create more time by opening a browser or something that should be accessible.

And then I think AI will be very helpful.

Thanks that we are we are on this journey now. Certainly from EULA twenty twenty four, it seems like the community is taking this on. We're not there yet, but I certainly I can see the direction of travel being there that will get better in terms of stratifying our patients to hopefully achieve one day, which is our great aim of personalized medicine. So thank you very much, Professor Hugo for your time today, and, hope you enjoy the rest of the EULA.

Thank you very much. Thank you.