Machine learning model for predicting outcomes of biologic therapy in psoriasis

To the Editor: The long-term treatment of psoriasis with biologics is associated with a gradual decrease of efficacy, leading to treatment discontinuation in a significant proportion of patients.1Egeberg A. Ottosen M.B. Gniadecki R. et al.Safety, efficacy and drug survival of biologics and biosimilars for moderate-to-severe plaque psoriasis.Br J Dermatol. 2018; 178: 509-519https://doi.org/10.1111/bjd.16102Google Scholar, 2Sbidian E. Mezzarobba M. Weill A. Coste J. Rudant J. Persistence of treatment with biologics for patients with psoriasis: a real-world analysis of 16 545 biologic-naïve patients from the French National Health Insurance database (SNIIRAM).Br J Dermatol. 2019; 180: 86-93https://doi.org/10.1111/bjd.16809Google Scholar, 3Yiu Z.Z.N. Mason K.J. Hampton P.J. et al.Drug survival of adalimumab, ustekinumab and secukinumab in patients with psoriasis: a prospective cohort study from the British Association of Dermatologists Biologics and Immunomodulators Register (BADBIR).Br J Dermatol. 2020; 183: 294-302https://doi.org/10.1111/bjd.18981Google Scholar As such, prognostic tools might be useful in optimizing long-term outcomes of biologic treatment. In this study, we compared the accuracy of a traditional statistical risk factor-based model versus machine learning (ML) in predicting the 5-year probability of biologic drug discontinuation. The Danish registry, Dermbio, comprising 6172 treatment series (ie, periods of continuous use of a biologic agent) in 3388 uncommon patients, was used as a data source.4Gniadecki R. Kragballe K. Dam T.N. Skov L. Comparison of drug survival rates for adalimumab, etanercept and infliximab in patients with psoriasis vulgaris.Br J Dermatol. 2011; 164: 1091-1096https://doi.org/10.1111/j.1365-2133.2011.10213.xGoogle Scholar,5Gniadecki R. Bang B. Bryld L.E. Iversen L. Lasthein S. Skov L. Comparison of long-term drug survival and safety of biologic agents in patients with psoriasis vulgaris.Br J Dermatol. 2015; 172: 244-252https://doi.org/10.1111/bjd.13343Google Scholar Biologics included in our study were adalimumab, etanercept, guselkumab, infliximab, ixekizumab, secukinumab, and ustekinumab. Variables analyzed included age, sex, body mass index, age at diagnosis, age at first biologic, prior exposure to biologic, concurrent psoriatic arthritis, concurrent methotrexate, presence of comorbidities, baseline Psoriasis Area and Severity Index (PASI), and baseline Dermatology Life Quality Index (DLQI). Hazard ratios were computed for all available predictive factors using Cox regression analysis with drug discontinuation being the outcome and adalimumab as the reference value. ML models, including Generalized Linear Model, Naive Bayes, Deep Learning, Decision Tree, Random Forest, and Gradient Boosted Trees, were trained using the 5-fold cross-validation technique and the abovementioned clinical characteristics. Model performance was assessed using the area under the receiver operating characteristic (AUROC) curve. Compared with adalimumab as a benchmark, the biologics associated with the lowest risk of discontinuation were ustekinumab and ixekizumab, whereas etanercept was found to have the highest risk of biologic discontinuation. Prior exposure to biologic therapy and patient sex were also significant variables, with drug survival being the longest in male patients with no prior biologic exposure. Weight and baseline PASI score were statistically significant predictors, but their contribution was negligible in comparison with the 3 aforementioned predictors (see Supplementary Material, available via Mendeley at https://doi.org/10.17632/cnkw7cdg4k.1). The nomogram in Fig 1 allows one to calculate the probability of drug discontinuation based on the Cox regression-derived predictors; however, the AUROC curve was 0.61 (Fig 2), which indicates a low discriminatory value.Fig 2Receiver operating characteristic (ROC) curves reflecting the ability to predict the 5-year risk of discontinuation. Curves indicate performance of the best-performing machine learning algorithm, Gradient Boosted Trees (GBT) compared to Cox regression analysis, with GBT being superior.View Large Image Figure ViewerDownload Hi-res image Download (PPT) All ML algorithms predicted the likelihood of discontinuation of a biologic within 5 years with high accuracy, ranging from 65.3% for Naive Bayes to 77.5% for Gradient Boosted Trees (see Supplementary Material). Thus, the most efficient ML algorithm was able to predict treatment outcome with less than 23% classification error, only utilizing basic patient information routinely available to every clinician. Similar to Cox regression analysis, the most important predictive parameters were the biologic drug (weight 0.247), patient sex (weight 0.076), and patient body weight (weight 0.069). The AUROC curve measured for the Gradient Boosted Trees was 0.85, which is an indicator of excellent performance of the algorithm (Fig 2). Ultimately, an ML-based approach, more so than a traditional statistical model, accurately predicted the risk of discontinuation of biologic therapy within 5 years of treatment based on simple patient variables available to dermatologists in clinical practice. The results of our study could fulfill an unmet need to predict the long-term effectiveness of biologics in patients with psoriasis. Dr Dalager has served on advisory boards with AbbVie, LEO Pharma, and Eli Lilly and has received honoraria as a consultant from Eli Lilly. Dr Skov has received research funding from Novartis, Bristol-Myers Squibb, AbbVie, Janssen Pharmaceuticals, the Danish National Psoriasis Foundation, the LEO Foundation, and the Kgl Hofbundtmager Aage Bang Foundation and honoraria as a consultant and/or speaker for AbbVie, Eli Lilly, Novartis, Pfizer, LEO Pharma, Janssen, UCB, Almirall, Bristol-Myers Squibb, Boehringer Ingelheim, and Sanofi. She has served as an investigator for AbbVie, Pfizer, Sanofi, Janssen, Boehringer Ingelheim, AstraZeneca, Eli Lilly, Novartis, Regeneron, Galderma, and LEO Pharma. Dr Thomsen has been a paid speaker for AbbVie, Eli Lilly, Novartis, Sanofi, Pierre Fabre, GSK, and LEO Pharma and has served on advisory boards with AbbVie, Eli Lilly, Janssen, Novartis, Roche, Sanofi, UCB, and LEO Pharma. He has served as an investigator for AbbVie, AstraZeneca, Boehringer, UCB, CSL, and Novartis and received research grants from AbbVie, Novartis, Sanofi, and UCB. Dr Gniadecki reports carrying out clinical trials for AbbVie and Janssen and has received honoraria as a consultant and/or speaker from AbbVie, Bausch Health, Eli Lilly, Janssen, Mallincrodt, Novartis, and Sanofi. The authors do not have equity in pharmaceutical companies. Drs Du, Ali, Ajgeiy, Dam, Egebjerg, Nissen, and Emam have no conflicts of interest to declare.