Predicting the Best Biologic Agents for Psoriasis

AI Artificial intelligence machine learning
The best biologic therapy for individual patients with psoriasis is found using predictive statistical and machine learning models.

Machine-learning and predictive models were able to predict which biologic agents led to psoriasis treatment duration of 3 years without discontinuation, according to findings from a population-based cohort study published in JAMA Dermatology.

Investigators obtained data on patients receiving biologic treatment for psoriasis via DERMBIO, a Danish national patient registry. The study included data on those 18 years of age and older at the time of treatment initiation and who had received either adalimumab, etanercept, infliximab, certolizumab pegol, ustekinumab, secukinumab, ixekinumab, or brodalumab. Follow-up time was either 1, 2, or 3 years, respectively. Treatments discontinued due to remission were considered successful regardless of the treatment duration. Investigators included information about comediation within the previous 6 months, comorbidities within the previous 5 years, and blood test results within the previous 3 months.

Through clustering performed on a small subset of variables, including Psoriasis Area and Severity Index (PASI) score and baseline Dermatology Life Quality Index (DLQI), 7 different patient types/clusters of clinical relevance were identified. Investigators used the algorithm “k-prototypes” which can appropriately handle a mix of continuous and categorical variables. Success rates of each biologic agent were computed for the clusters assuming a success criterion of 2 years for all treatment series and separately for bionaive and non-bionaive treatment series. Investigators used logistic regression models as benchmark and compared them to gradient-boosted decision trees.  Performance was reported as accuracies, top 2 accuracies, top 3 accuracies, and receiver operating characteristics area under the curve (ROC AUCs) when applicable. There were 3 different approaches for imputations with model performances reported for each: patient characteristics as mean/median with standard deviations and trajectories of PASI and DLQI scores plotted by interpolating between dermatology different visits for each treatment series; binary prognostic models to predict discontinuation within 1, 2 and 3 years, respectively; and multiclass models to classify patients based on the specific biologic therapy possibly resulting in successful treatment.

For success criteria of 1, 2, and 3 years, the study encompassed 4762, 4041, and 3452 treatment series distributed among 2668, 2340, and 2034 patients, respectively. More than half of patients in treatment series with a success criterion of 2 years were men (61.4%), and the mean age at diagnosis and prescription date were 24.7 years and 45.7 years, respectively. The mean baseline PASI score was 9.5.

For successful treatment series, the PASI score started around 10 and quickly decreased within the first 100 days, plateauing slightly below 2, researchers wrote. Unsuccessful treatment series started around 10 and decreased within the first 100 days, plateauing around 4 but increasing approaching 700 days of treatment. The same pattern was seen for DLQI scores.

The algorithm divided patient clusters into subgroups ranging from 356 to 699 participants. In cluster 6, characterized by female sex, absence of psoriatic arthritis, and high PASI and DLQI score, bionaive patients appeared to have lower success rates when treated with infliximab compared to other drugs. Nonbionaive patients in cluster 2, characterized by male sex, absence of psoriatic arthritis, and high PASI and DLQI score, and cluster 3, characterized by male sex, absence of psoriatic arthritis, young age at diagnosis, and low age at biologic initiation, had higher success rates with infliximab compared with etanercept, ustekinumab or secukinumab. However, bionaive patients in cluster 2 had lower success with adalimumab and more success with secukinumab. Bionaive patients in cluster 3 had higher rates of success with either infliximab, secukinumab, or ustekinumab and lower success with etanercept. Overall, patients in cluster 0 which was characterized by male sex and concomitant psoriatic arthritis had the highest success rate when treated with secukinumab compared with other drugs.

In models predicting the success or failure of treatments for all therapies combined, performance improved slightly with less strict success criteria. In binary models, accuracies increased with less strict success criterion, and gradient boost involving ustekinumab generally performed better than models involving other treatments. Models involving infliximab performed poorly, especially with a success criterion of 2 years. In multiclass models predicting targets and drugs, gradient boost had an accuracy of 63.6% and a top 2 accuracy of 95.9% for a success criterion of 3 years. That is, models were able to predict the correct target for individual patients in 63.9% of the cases and the correct target in top 2 in 95.9% (which is equivalent to excluding 1 target), which is the equivalent to excluding 1 target. When predicting specific agents, gradient boost has a top 2 accuracy of 48.5%, top 3 accuracy of 77.6%, and 88.9% with a success criterion of 3 years. Gradient boost performed significantly better than logistic regression for all success criteria when predicting drug/target, except when predicting a target with a success criterion of 2 years.

Variables with the highest discriminative power included previous history of biologic treatment, age at diagnosis and prescription date, weight/BMI, baseline PASI and DLQI scores, and status of psoriatic arthritis diagnosis.

The study was limited by large amounts of missing data for some variables and drugs which could not be included in the models. In addition, since patients were only given a single biologic agent at a time, models could not predict whether another drug given at that point in time would have led to success.

“These findings could potentially be useful for clinicians in future treatment decision processes for choosing of appropriate biologic therapy for patients with psoriasis,” the study authors wrote.

Disclosure: Several study authors declared affiliations with biotech, pharmaceutical, and/or device companies. Please see the original reference for a full list of authors’ disclosures. 


Nielsen ML, Petersen TC, Maul JT, et al. Multivariable predictive models to identify the optimal biologic therapy for treatment of patients with psoriasis at the individual level. JAMA Dermatol. Published online August 17, 2022. doi:10.1001/jamadermatol.2022.3171