Understanding patterns of rheumatic disease trajectory and treatment response
There are number of rheumatic diseases that have similar and overlapping sets of symptoms that make them challenging to diagnose correctly in the first instance. Additionally, patients can have different disease trajectories and respond differently to different drug therapies. Machine learning data analysis has helped to define different subgroups of patients either at baseline or from longitudinal follow-up data, and associate them with different disease trajectories, as well as different responses to medication. The aim of this work is to be able to better understand the nature of the disease for the individual and to make personally tailored treatment decisions earlier. CLI chatted with Professor Nophar Geifman (University of Surrey, UK) to learn more about her team’s work.
There are a number of rheumatic diseases, can you begin by telling us something about these different diseases and their pathogenesis?
There’s a large number of what we would call rheumatic conditions or diseases, which include a wide range of different and distinct conditions that manifest with an overlap of symptoms, but each with their own specific nuances. Generally speaking, they all affect the joints, muscles, bones, tendons, ligaments or other connective tissue, but some of them also can involve internal organs as well, such as the heart, lungs, kidneys, skin, and nervous system. So they really can be quite systemic conditions.
These conditions can broadly be split into two groups. There are the inflammatory autoimmune diseases, where the immune system plays a critical role in the pathogenesis of those conditions. And then there are some non-inflammatory degenerative conditions that have other aspects that affect the pathogenesis. These could be mechanical or metabolic or even injury-related conditions. It’s perhaps quite a coarse way of splitting the diseases into groups, but historically, in my group, we have tended to focus on the inflammatory autoimmune conditions, also known as immune mediated inflammatory diseases (IMIDs). This includes several of these rheumatic conditions that have an inflammatory aspect to them – such as rheumatoid arthritis (RA). In general (and it is a generalization because within each condition there are outlier patients that don’t manifest with the classical symptoms), the joints are affected but also the blood vessels as well as synovial inflammation, and there’s a key prominent autoimmune aspect to that condition. The prevalence in the UK is approximately 1%.
Then, there are other rarer conditions such as lupus [systemic lupus erythematosus (SLE)], which affects approximately 100 people per 100 000 in the UK, with a higher proportion of patients being female.
Other conditions such as psoriatic arthritis (PsA), still see debate about whether the definition of the disease is accurate or not. PsA is a combination of psoriasis, which is quite prevalent, and arthritis. Yet, it’s different from rheumatoid arthritis and it’s different from psoriasis, but it combines manifestations from both. Patients have the skin manifestations from psoriasis but the joints are also involved.
We’re also interested in juvenile idiopathic arthritis (JIA), which occurs in children and young adults – the definition includes people up to the end of their teens and into the start of their twenties. This condition has similarities with adult arthritis; however, interestingly, some patients will outgrow the condition and achieve complete remission as they turn into adults, whereas others will continue to have active disease.
So that’s just a few of the rheumatic diseases – the list is quite long once you start to delve into it.
Pathogenesis
We don’t necessarily know exactly what causes these diseases. We understand the mechanism a little bit better for some than others. We’ll focus here on the conditions that have an inflammatory–immune aspect. Broadly speaking, there are different parts of the immune system that end up out of control and start misbehaving. Why? is a good question. Professor Luke O’Neill in a recent Zoe podcast1 talks about how we all have some level of inflammation going on in our body, and that’s potentially a good thing when it’s at a very low level, because it means our immune system is primed to react when needed. However, we’re not entirely sure why some people overreact and why sometimes that overreaction turns into a condition/disease. There’s not a single cause, but there are genetic factors and thus inherited factors that affect susceptibility, and lifestyle factors can also affect the risk of developing some of these conditions. So, although we don’t know specifically how all these diseases arise it is currently a very active area of investigation.
How is diagnosis normally achieved and what are the challenges with diagnosis?
Similar and overlapping symptoms
Diagnosis is challenging because the symptoms tend to be similar across many of these conditions which affect the joints, bones and connective tissue. There are similar symptoms of joint swelling and pain, stiffness, as well as more general symptoms such as fatigue and pain more generally.
However, clinicians start with a medical history and family history. Physical examination will be done, as well as imaging (X-radiography, ultrasound, MRIs) if needed, and a range of different blood tests.
Nonspecific biomarkers
The blood tests involve testing a range of biomarkers, such as C-reactive protein (CRP), erythrocyte sedimentation rate (ESR) for inflammation; antinuclear antibody test (ANA) for autoimmune diseases in general; rheumatoid factor (RF) and anti-cyclic citrullinated peptide (anti-CCP) for rheumatoid arthritis; and-dsDNA for SLE. However, many of these biomarkers are not great. CRP and ESR indicate inflammation, which can come and go for different reasons and an elevated level is not necessarily specifically indicative of these diseases. RF is relatively specific for RA, although some people will have elevated levels without having the disease and it doesn’t identify all patients with RA; however, if a patient tests positive for RF it is more likely that they do have RA as opposed to one of the other conditions.
Delayed diagnosis and misdiagnosis
So, although the biochemical tests are not definitive, they do play a part in the disease measure, which takes into account different aspects of the disease. For instance, there could be a disease score that can include a physician measurement, for example in psoriasis this would include the area of the skin that’s affected with plaques and how flakey/thick/peeling they are, which can be combined with the biological markers from the blood test results. In RA and JIA, the patient reported measure would also be taken into account: the patient would say how much stiffness, pain, etc. they have, in addition to the clinician’s measurements of how swollen their joints are. Depending on the healthcare system, different scores or combinations of these scores are used to say that this patient has this disease with this degree of severity. However, diagnosis often tends to be a process of elimination and so can take a long time depending on condition. A diagnosis of SLE, because it’s rarer, tends to take longer as it happens after the other, more common conditions have been excluded. Unless someone presents with the classical SLE symptoms (e.g. butterfly rash), it’s harder to reach that diagnosis. This means that patients do tend to go a significant period of time before they have a diagnosis and so treatment is delayed, which of course in progressive diseases like this is unhelpful.
Also, patients do not necessarily present with the classical symptoms for a disease – for example, PsA patients can present with classical psoriasis symptoms and so end up with a misdiagnosis. We’re investigating those cases of patients with rarer diseases who end up potentially with the wrong diagnosis because they either didn’t present with the classical symptoms or they presented with symptoms that the physician then thought meant that they fitted into a different diagnosis box. Often we find that looking at patients longitudinally rather than as a single snapshot means that we’re getting a clearer picture of how the diseases change over time and the markers for that are clearer, more precise. However, this is currently not yet implemented as a diagnostic tool.
Effect on wellbeing
The conditions themselves, along with the challenges of perhaps a delayed diagnosis or misdiagnosis, have a knock-on effect on the patients’ wellbeing [Low et al., 2024]. Patients with these diseases can’t get around as easily and we’re seeing associations with other mental health conditions such as depression, and so on. So, yes, when a patient presents, it’s not always as straightforward to assess fully what’s actually happening – diagnosis can be difficult.
1 “How to eat to lower inflammation: The top 5 foods to boost energy, stop pain, and live longer with Prof. Luke O’Neill”; October 2026 (https://zoe.com/learn/how-to-eat-to-lower-inflammation?lead_id=211cb15d-1016-4ce5-ad30-f6138692f231)
Autoimmune activity in rheumatoid arthritis
How are these diseases normally treated?
Again, there are a lot of commonalities across these different conditions with an escalating approach to treatment, although there are nuances and differences between different conditions and how that treatment is provided. Some treatments will be administered orally, some will be injections directly into the joint, or some will be intravenous, depending on the condition and on the medication. Generally, for some of these conditions, the patient will usually start off with a non-steroid anti-inflammatory drug (NSAID) treatment, starting with ibuprofen, naproxen, etc. Then, in some, there will be corticosteroid injections into the articular space, which is unpleasant, but not something that has to be done every day but every few months. That can help a lot with the symptoms and to slightly slow down local disease progression. The more common, well-known medication is methotrexate, an immunosuppressant disease-modifying antirheumatic drug (DMARD). This helps slow down the progression of the disease by reducing the damage caused by the immune system on whatever tissue the immune system is attacking. It is considered to be quite effective, but it has, as with probably every treatment, its downsides – some patients suffer from significant side effects and adverse events that affect their wellbeing, and so it’s not optimal for everyone – and it’s not ideal to be on immunosuppressant medication for the long term.
In many of these conditions, arthritis, JIA, Psoriasis, SLE, once a patient is either not showing any response to DMARDs or they’re suffering from side effects or their disease is progressing to a point where it’s quite severe, a different intervention is needed. The National Institute for Health and Care Excellence (NICE) has guidelines for each condition regarding the cut-off, but once a patient has reached that level, they’ll move on to a class of drugs called biologics (immunotherapy), which are antibody-based drugs that inhibit certain aspects of the immune system. Anti-TNF antibody drugs are probably the most common ones that are used, for example, adalimumab, etanercept, and infliximab. There are also other biologic drugs that target other aspects of the immune system, including interleukins, T- and B-cells and Janus kinase (JAK). The biologics tend to be very effective, life changing even, for many patients.
What are the challenges of treating these diseases?
Variable treatment response
As mentioned above, the different therapies have different challenges – not everyone can tolerate injections into the joint and some people do have intolerable side effects from the medications. The biologic therapies are very effective in many people but not everyone. For some reason (and I don’t think for any condition do we know what the reason is), some individuals have poor response or don’t respond at all, and some improve and then have a relapse and the disease will flare up again.
Psoriasis
Clinical decision-making algorithms
Biologics, however, are expensive, and in the UK, because we have a National Health Service, there has to be a system that allows a careful decision to be made about who should receive those therapies to optimize the use of resources. For example, with psoriasis, in order for a clinician to prescribe biologic therapy, the patient has to reach quite a high threshold of disease severity – the NICE guidelines require a psoriasis area and severity index (PASI) relatively high score of 10 or greater (along with a dermatology life quality index [DLQI] score greater than 10) to qualify for biologic therapy in adults with moderate-to-severe plaque psoriasis. However, we knew that clinicians were seeing a percentage of patients who are not responding to these therapies. Although clinicians need these guidelines for making decisions, the reality is, in diseases that are as complex as arthritis, lupus, psoriasis, those thresholds can be quite arbitrary, based on perhaps old, anecdotal evidence: they’re helpful in clinic but they’re not as helpful if you’re trying to understand the disease or improve diagnoses. For example, with patients on biological therapy, the cut-off PASI score is 10; if one has a score of 10.2, they are classed as a non-responder, whereas another, with a score of 9.7 can be classed as a responder because their disease severity has been lowered. However, are those two points really that different? Is that marginal difference really biologically meaningful? The answer is probably not. We apply arbitrary thresholds because we have to, because we need a way to make decisions. But this leads to us to dichotomize disease: they have the disease or they don’t have the disease; they respond to medications or they don’t respond to medication. Additionally, clinicians were not really looking at the patient’s disease trajectory and computationally including these measurements over time in the thresholds that were established and the decisions that were made.
Evidence-based treatment response groups
So when biologic therapies were relatively new, in the UK, registries were formed so that data from patients who receive a biologic are collected centrally. Initially this was to watch for any indication that treating people with these drugs might cause long-term adverse outcomes, such as in increase in cancer rates. It turned out not to, but this meant that we had disease-focused data sets with longitudinal, very detailed data collected on these patients for these conditions. We were given access to the psoriasis data set (BADBIR; https://www.badbir.org) to see if we could identify clinical phenotypes that would allow us to examine response trajectories of approximately 3500 psoriasis patients treated with biologics [Geifman et al., 2021]. The messy, real-world data where each patient’s response trajectory is represented as a line looked like a dense cloud of lines going in all different directions and was computationally challenging to deal with. However, we found an interesting semi-statistical, semi-machine learning, regression-based, method that uses a latent approach to find groupings within these longitudinal patterns. We were trying to see if we could find groups of patients whose disease/treatment response followed a similar trajectory over time. We identified a group that responded really well as well as a group that didn’t respond well, and, two additional groups with intermediate response trajectories. Then, when we looked at their age and the different symptoms that they present with at baseline, we found associations between clinical phenotypes that differed between the groups, and they generally made sense clinically. For example, patients with palm involvement tended to have lower response and when clinicians looked at our data they confirmed that it matched what they saw in clinic, but just formalized a bit more.
Additionally, we wanted to see if we could find a completely objective marker that we can assess when a patient presents in clinic for the first time and is diagnosed with psoriasis that will indicate if they are going to respond to a biologic or not, or maybe they’ll respond a little bit. Indeed, we found human leukocyte antigen (HLA) biomarkers that helped to distinguish between the different trajectory-based subgroup of responders versus non-responders. Of course, these genetic markers are fixed and don’t change over time.
We also examined transcriptomic data in a smaller number of patients, and found a set of transcriptomic markers that differentiate in particular between the good responders and the non-responders. The ones in the middle that showed some response were a little bit harder to differentiate. That obviously has potential for greater impact because that could become a test for when a patient’s disease reaches that level of severity where they require more specific treatment, allowing an evidence-informed decision to be made about giving a patient a drug and knowing that they will respond well, or not giving it to a patient because it’s less likely to help them.
Methotrexate is often a first-line treatment for rheumatoid diseases
JIA
We have used a similar approach across the different diseases but JIA has been a very interesting, slightly different question. Firstly, again, we knew that with JIA, some children respond to treatment whereas other don’t. Some children can tolerate methotrexate less well or have adverse response to it, especially long term. However, with pediatric patients, there are additional challenges in that, until the child’s old enough to articulate exactly what hurts, where and when, we are communicating with the patient’s parent (or proxy) rather than the patient themselves, which risks introducing a slight bias. So it’s usually the patient’s parent who will says things like: My child isn’t eating as much; They are not sleeping as well; or It hurts then when they walk. And, of course, children have different abilities at different ages. So, there are different aspects that affect our interaction with children that can make it difficult to capture what’s happening with them.
Different clusters of disease at baseline and different disease trajectory groups
We very successfully described a lot of the issues that affect children [Shoop-Worrall et al., 2020]. We used the same techniques as in the psoriasis study and identified five clusters of JIA at baseline and six disease trajectory groups using longitudinal follow-up data, based on physician and patient assessment of disease. One of our first key findings was that there was a subgroup of children whose disease manifestation using more objective terms is very different to their perceived or experienced disease. For example, their inflammatory markers were going down and it looks as if they’re getting better – their disease should be better but actually they’re feeling worse. So it’s very important for the clinician to listen carefully to the patient because sometimes the markers that we think are specific, objective, and even the doctor’s observations, do not necessarily fully capture what’s going on with the patient.
Response to methotrexate treatment
We were fortunate that, because methotrexate is the drug used in the first instance, there’s quite a lot of children treated with methotrexate and we had access to data from four UK cohorts. We applied the same type of approach looking at these trajectories to find different sub-types of response to methotrexate [Shoop-Worrall et al., 2024]. And, again, we found six different groups with some children responding a lot better, some children responding a lot less, and nuances in between. The approach that we took with the children, because of what I mentioned earlier, was to simultaneously model different aspects of the disease, so that we capture different disease aspects that are measured by different people, or are more/less objective, and where children might have less or more pain, but less or more inflammation at the same time. We found that factors associated with the different response groups included ethnicity, The International League of Associations for Rheumatology (ILAR) category, age, parental global evaluation (PGE), and ESR scores at commencement of methotrexate therapy. This was quite encouraging because it echoed what we saw in the more general JIA population mentioned above, but also pinpointed those groups that do better or have different kinds of manifestations earlier on.
Response to etanercept therapy
Following our study on patient response to methotrexate, we again used a similar approach to study patients treated with etanercept [Shoop-Worrall et al., 2026]. Of course, these are patients who have already tried other therapies, so while it may be the patient’s first biologic therapy, it’s most likely not to be their first line of treatment. Again, we also looked at how patterns longitudinally mirror or match up to the different groups of patients at a single time point at presentation. So, can you actually identify groups of children that are with different outcomes long term by looking only at the beginning of that trajectory, the start of their disease? And the answer is an encouraging tentative yes.
Implications
However, the more data you have, the longer you follow patients, the better the computational methods work and, therefore, we tend to have better outcomes. But we want to not have to treat a patient for three years unsuccessfully, only then to know what the right way to treat them is – we want to move as early as possible. And the same goes for diagnosis. Because these diseases are progressive and degenerative, it would be ideal to make a diagnosis before symptoms begin, but that’s quite difficult. In order to do that kind of research you’d essentially have to monitor everyone for everything until they end up with a disease.
Most of this work was done before generative and agentic AI was making the headlines, so we did some work with children and practitioners to see how acceptable they would find these modelling approaches. As a data scientist, it’s easy to work in a bubble and not really engage with patients or people who are going to be affected: we publish a paper; here are six trajectories following methotrexate, go do something with it. But, we carried out workshops with young people, and practitioners, showing them the modelling and asking: Does this resonate? Does this make sense? Do you agree? Do you not agree? Would you change anything? Do you find this beneficial? [Shoop-Worrall et al., 2021]. Their feedback was hugely beneficial to the research directions we take. With JIA, we really have tackled a lot of that system with different approaches.
Psoriasis
How might diagnosis and treatment be improved?
The ultimate aim is to give patients the correct diagnosis at first presentation and then also to be able to say that this patient will or will not benefit from a given treatment, i.e. the right treatment for the right patient at the right time. The ability to do this is also very disease specific. For the SLE study, we took a slightly different approach. We carried out a proteomics analysis on over 200 patients – which, although it sounds like a small number, is quite a lot because SLE is so rare. Following the mass spectrometry proteomic analysis on these patients, we used another clustering approach to see if different groups of patients could be identified. And we did find that that a snapshot of blood biomarker proteins can demonstrate a strong association with different clinical outcomes and characteristics. I think that’s also where we want to go with other conditions as well. Ideally, there would be a biomarker test that is specific and precise. This is very likely not going to be a single biomarker but a panel, as these conditions are so complex that you’re never going to capture it with a single protein. It will be some combination of different proteins or even different types of molecules, such as proteins and RNA combined together that will give that accuracy that is needed. So, ideally, there would be a biomarker panel that could differentiate between the top five conditions that have that similar symptomology and then be able to inform therapeutic decision-making. This would be good for patients and also for the healthcare system, as these are important conditions – they affect a lot of people, they can be very debilitating and treatment can be expensive. I’ve been privileged to work with many patients from across these different conditions, and you see how it affects their daily lives, how difficult it is – don’t take your joints for granted!
Additionally, better understanding of the mechanisms and the causes of the diseases and, therefore, potentially being able to prevent them or at least mitigate their effects would be the ultimate long-term goal. We know that lifestyle factors can influence the manifestation and severity of almost all these conditions, so a better understanding of the causal mechanisms will mean that you can tailor treatment not just in terms of drugs but also in changes to lifestyle factors, and inform patients better. What I can say is that we are in a far better position concerning understanding patients’ cases individually and thereby helping promote wellbeing now than we were 10 years ago.
Final comments
This is a work in progress, and we’ve been very fortunate to work with such fantastic groups of researchers and patients. Especially with conditions like these, that are so complex, it is vital that we have collaborations across different stakeholders and are not working in isolation. We have been privileged to have these collaborations across the disease ecosystem for many of these conditions, and we are always very grateful for patients allowing us to use their data because it’s not trivial, especially in today’s cyber climate – we would not be making any progress if we didn’t have access to that kind of data. It really is team science in the purest meaning and that’s really needed for these and many other conditions.
Bibliography
Juvenile idiopathic arthritis (JIA)
- Shoop-Worrall SJW, Hyrich KL, Wedderburn LR et al.; CAPS the CLUSTER Consortium. Patient-reported wellbeing and clinical disease measures over time captured by multivariate trajectories of disease activity in individuals with juvenile idiopathic arthritis in the UK: a multicentre prospective longitudinal study. Lancet Rheumatol. 2020;3(2):e111–e121. PMID: 33554133 (https://doi.org/10.1016/s2665-9913(20)30269-1). Erratum in: Lancet Rheumatol. 2021;3(4):e250. PMID: 33860279 (https://doi.org/10.1016/s2665-9913(21)00082-5).
- Shoop-Worrall SJW, Cresswell K, Bolger I et al.; Members of the CLUSTER consortium. Nothing about us without us: involving patient collaborators for machine learning applications in rheumatology. Ann Rheum Dis. 2021;80(12):1505–1510. PMID: 34226185 (https://doi.org/10.1136/annrheumdis-2021-220454).
- Shoop-Worrall SJW, Lawson-Tovey S, Wedderburn LR et al.; CLUSTER Consortium. Towards stratified treatment of JIA: machine learning identifies subtypes in response to methotrexate from four UK cohorts. EBioMedicine. 2024;100:104946. PMID: 38194741
(https://doi.org/10.1016/j.ebiom.2023.104946). - Shoop-Worrall SJW, Lawson-Tovey S, Wedderburn LR et al.; UK JIA Biologics Register Investigators, CAPS Investigators, CHARMS Investigators, CLUSTER Consortium. A stratified approach to juvenile idiopathic arthritis: leveraging machine learning to uncover subgroups of patients following treatment with etanercept. Rheumatol Adv Pract. 2026;10(3):rkag088. PMID: 42583531 (https://doi.org/10.1093/rap/rkag088).
Rheumatoid arthritis (RA)
- Dagliati A, Plant D, Nair N et al.; BRAGGSS Study Group. Latent class trajectory modeling of 2-component disease activity score in 28 joints identifies multiple rheumatoid arthritis phenotypes of response to biologic disease-modifying antirheumatic drugs. Arthritis Rheumatol. 2020;72(10):1632–1642. PMID: 32475078 (https://doi.org/10.1002/art.41379).
Psoriasis
- Geifman N, Azadbakht N, Zeng J et al.; BADBIR Study Group, on behalf of the PSORT Consortium. Defining trajectories of response in patients with psoriasis treated with biologic therapies. Br J Dermatol. 2021;185(4):825–835. PMID: 33829489 (https://doi.org/10.1111/bjd.20140).
- Low JM, Hyrich KL, Ciurtin C et al.; CAPS Principal Investigators. The impact of psoriasis on wellbeing and clinical outcomes in juvenile psoriatic arthritis. Rheumatology (Oxford). 2024;63(5):1273–1280. PMID: 37467079 (https://doi.org/10.1093/rheumatology/kead370).
Systemic lupus erythematosus (SLE)
- Reynolds JA, Prattley J, Geifman N et al.; MASTERPLANS Consortium. Distinct patterns of disease activity over time in patients with active SLE revealed using latent class trajectory models. Arthritis Res Ther. 2021;23(1):203. PMID: 34321096 (https://doi.org/10.1186/s13075-021-02584-x).
- Su KYC, Reynolds JA, Reed R et al.; the MASTERPLANS and BILAG-BR consortia. Proteomic analysis identifies subgroups of patients with active systemic lupus erythematosus. Clin Proteomics. 2023;20(1):29. PMID: 37516862 (https://doi.org/10.1186/s12014-023-09420-1).
The interviewee
Professor Nophar Geifman* M Med Sc, PhD, FHEA
Professor of Health and Biomedical Informatics, and Director of Informatics, the Veterinary Health Innovation Engine (vHIVE)
School of Health Sciences, Faculty of Health and Medical Sciences, University of Surrey, Guildford, UK
Email: n.geifman@surrey.ac.uk
*Professor Geifman is a data scientist not a rheumatologist.