Data‐driven prediction of biologic treatment responses in psoriasis: steps towards precision medicine

Short abstract

Linked Article: Geifman et al. Br J Dermatol 2021; 185:825–835.

In this issue of the BJD, Geifman et al. defined subgroups or ‘trajectories’ of patients with psoriasis with similar patterns of disease severity (Psoriasis Area and Severity Index, PASI). The patients were studied over time using a data‐driven latent class mixed‐modelling approach. ¹ This modelling approach can be used to divide a heterogeneous population into a few homogeneous groups or ‘trajectories’. As such, patient characteristics in specific trajectories can be used to predict health outcomes such as treatment success. ² , ³ With the increase in (expensive) treatment options for psoriasis, but still a substantial number of nonresponders per biologic, these analyses are very welcome. In this study, data of patients who were treated with various biologics were included. Four PASI trajectories were identified, with differences in clinical characteristics such as body mass index, baseline PASI, psoriasis subtype and the specific biologics.

In pharmacoepidemiological studies, outcomes, exposures and presence of patient characteristics are often included in analyses according to a binary approach (present vs. not present). However, this might not accurately reflect the real‐world situation, where outcomes, exposures and characteristics such as comorbidities can vary over time. Alternative modelling methods, such as trajectory modelling techniques, are able to summarize complex individual‐level medication‐utilization trajectories or time‐varying exposures. Such models result in several groups of patients within a given population who share similar patterns of characteristics over time. ³ Within the study of Geifman et al., an unsupervised data‐driven approach was used to identify subgroups of patients with similar patterns of PASI scores over time. A strength of this approach is the possibility of handling large datasets and to be able to separate them into groups without bias of pre‐existing knowledge. A possible disadvantage is that such unsupervised methods are at risk of overextraction or identifying groups that are not ‘true’. In addition, different methods for clustering groups might produce different subgroups. ⁴ Reassuringly, the authors conducted the analysis in two independent cohorts, which showed overall similarities.

In the study of Geifman et al., different biologics were grouped due to data availability. ¹ As such, the trajectories were not biologic specific. However, a sensitivity analysis on adalimumab alone suggested that the identified trajectories were generic instead of treatment specific. From a clinical perspective, it would be interesting in future studies to define biologic‐specific trajectories, as different biologics might provide different treatment responses over time. In addition, Geifman et al. indicated that in future studies additional molecular and pharmacological data could add to definitions of subgroups and eventually lead to a more accurate prediction of treatment responses. Treatment‐specific trajectories and inclusion of biomarkers in these models would lift precision medicine to a higher level.

In conclusion, the ambitious data‐driven approach used in the study of Geifman et al. is a promising first step in the use and evaluation of big (observational) data for the prediction of psoriasis treatment responses and treatment optimization in the future. Back‐translation and implementation of such results into clinical practice will be an important future challenge. The ongoing rise of big data brings multiple opportunities, which require us to embrace an intensive collaboration between data scientists and clinicians.

Author Contribution

Lara Sophie van der Schoot: Writing‐original draft (equal); Writing‐review & editing (equal). Juul M.P.A. van den Reek: Writing‐original draft (equal); Writing‐review & editing (equal).