JMIR Publications

ABSTRACT

Background:

Harnessing longitudinal data for time to event analysis can provide valuable insights into disease progression and help plan clinical interventions for individual patients, with the goal of improving clinical outcomes and quality of life. However, real-world clinical data is characterised by missingness, inconsistencies and heterogeneity, especially when datasets are aggregated from different sources.

Objective:

To address the challenges of missingness, inconsistency, and heterogeneity in multi-source data of degenerative disease, we propose a framework for explaining time-to-event predictions using multivariate longitudinal trajectories, applied to time-to-gastrostomy in patients with Amyotrophic Lateral Sclerosis (ALS).

Methods:

We analysed data from 8,586 ALS patients using a two-stage analytical approach. Joint latent class growth discrete-time survival analysis were used to identify data-driven reference trajectories of functional decline. New patient markers were mapped to these clusters using Fréchet distances. Three survival models (Cox PH, Cox XGBoost, and XGBoost Pseudo-Observation Regression) using baseline demographics and functional decline features were used to predict time-to-gastrostomy.

Results:

Distinct classes of functional decline revealed that rapid deterioration in bulbar and swallowing functions is the most critical indicator for intervention, reaching a 50% probability of gastrostomy within 16 to 18 months. Bulbar and swallow onset slopes were the primary predictors of time-to-gastrostomy. Predictive models utilizing the early "onset slopes" of functional decline outperformed those using baseline demographics alone, yielding a 0.044-0.069 increase in concordance index and decreasing median absolute error by 60-157 days compared to relying on diagnostic delay. The XGBoost MAEPO regression model utilizing onset slope was the best overall classifier, achieving a concordance index of 0.731 (IQR, 0.717-0.744) and a median absolute error of 218 days (IQR, 204-232). Additionally, all evaluated models comfortably outperformed a naïve classifier based on a 10% weight loss threshold.

Conclusions:

Our framework addresses clinical data heterogeneity through principled feature extraction and unsupervised trajectory mapping, translating individual predictions into interpretable clinical narratives that support timely gastrostomy decisions, and more generally time-to-intervention in degenerative diseases.