EczemaPred: A computational framework for personalised prediction of eczema severity dynamics

Abstract

Background: Atopic dermatitis (AD) is a chronic inflammatory skin disease leading to substantial quality of life impairment with heterogeneous treatment responses. People with AD would benefit from personalised treatment strategies, whose design requires predicting how AD severity evolves for each individual.

Objective: This study aims to develop a computational framework for personalised prediction of AD severity dynamics.

Methods: We introduced EczemaPred, a computational framework to predict patient-dependent dynamic evolution of AD severity using Bayesian state-space models that describe latent dynamics of AD severity items and how they are measured. We used EczemaPred to predict the dynamic evolution of validated patient-oriented scoring atopic dermatitis (PO-SCORAD) by combining predictions from the models for the nine severity items of PO-SCORAD (six intensity signs, extent of eczema, and two subjective symptoms). We validated this approach using longitudinal data from two independent studies: a published clinical study in which PO-SCORAD was measured twice weekly for 347 AD patients over 17 weeks, and another one in which PO-SCORAD was recorded daily by 16 AD patients for 12 weeks.

Results: EczemaPred achieved good performance for personalised predictions of PO-SCORAD and its severity items daily to weekly. EczemaPred outperformed standard time-series forecasting models such as a mixed effect autoregressive model. The uncertainty in predicting PO-SCORAD was mainly attributed to that in predicting intensity signs (75% of the overall uncertainty).

Conclusions: EczemaPred serves as a computational framework to make a personalised prediction of AD severity dynamics relevant to clinical practice. EczemaPred is available as an R package.

Keywords: Bayesian model; PO-SCORAD; atopic dermatitis; machine learning; prediction.

FIGURE 1

Example trajectories of PO‐SCORAD and its severity items for representative patients from datasets 1 (A) and 2 (B)

FIGURE 2

Model overview. (A) Bayesian state‐space models in EczemaPred. Each model describes the dynamics of a latent severity (white ovals) and the measurement of the latent severity to obtain the recorded severity (grey ovals). (B) Use of EczemaPred for SCORAD prediction. Predictions from nine models (coloured rectangles), each of which corresponds to one of the nine severity items for SCORAD, are aggregated to provide predictions for SCORAD. (C) Latent dynamics and measurement distributions for the three severity components of SCORAD

FIGURE 3

Predictive performance for 4‐day‐ahead forecasts by EczemaPred models (empty circles) and reference models (filled circles) measured by lpd (the higher, the better). EczemaPred models are a binomial Markov chain model (BinMC) for extent, an ordered logistic random walk model (OrderedRW) for intensity signs, and a binomial random walk model (BinRW) for subjective symptoms. Reference models include a uniform forecast (uniform), a historical forecast (historical), a random walk model (RW), and a Markov chain model (MC). The performance was calculated after training with approximately 80% of the data (77 days' data for dataset 1 and 65 days' data for dataset 2)

FIGURE 4

PO‐SCORAD prediction by EczemaPred for four representative patients from dataset 1 (A) and dataset 2 (B). Coloured ribbons correspond to stacked prediction intervals of highest density (darkest ribbon corresponds to the mode), and black dots represent the recorded PO‐SCORAD. The model is updated, and new predictions are issued every 4 days (vertical dashed lines)

FIGURE 5

Learning curves for 4‐day‐ahead forecasts of PO‐SCORAD evaluated by lpd (top) and accuracy (bottom) as a function of the number of training observations (training days), for datasets 1 (left) and 2 (right). EczemaPred models perform better than reference models, including an exponential smoothing model (Smoothing), a mixed effect autoregressive model (MixedAR), an autoregressive model (AR), a random walk model (RW), a historical forecast (historical), and a uniform forecast (uniform)