Time is of the essence: Temporal Modelling of Frailty Profiles in Routine Health Data for Equitable Risk Stratification

Background

Frailty (often defined as a reduced resilience to stressors) is conventionally quantified as a Frailty Index (FI), which compresses information about multiple health deficits into a single score [1]. While predictive this approach carries two main limitations. 

Firstly, the aggregation implies that deficits are interchangeable units, obscuring critical heterogeneity in the nature of the deficits and limiting the efficacy of clinical interventions. For example, a patient accumulating cognitive deficits differs mechanistically from one accumulating physical deficits even if they have the same (total) FI score: not only would they be susceptible to distinct cause-specific outcomes (e.g., cardiovascular vs. respiratory mortality), but may potentially respond differently to the same treatments simply because the FI is but a summary of a complex system. Secondly, FI is often treated as a cross-sectional metric, with an underlying assumption that once a deficit is encountered, it is irreversible. Even where temporal modelling is considered [2], it is at the level of counts, which fails to consider the clinical utility of predicting patient frailty trajectory and how intervention on one dimension could adversely affect another. 

Taken together, these two assumptions create a dual gap: we lack models that (1) resolve frailty into distinct deficit subtypes, and (2) map the temporal transitions of these subtypes simultaneously. Further treating these as separate problems misses the interaction: in fact, there is recent evidence of distinct deficit subtypes (e.g., cognitive, physical, social) following divergent temporal dynamics (some with reversibility) [3], although more research is required to characterize this. 

These gaps are not merely methodological; they also limit our understanding of health equality. Socioeconomic deprivation may not simply accelerate the rate of accumulation (temporal) but also alter the trajectory of specific deficit types (compositional). For example, deprivation might precipitate some deficits earlier than others, a nuance invisible to aggregate temporal models. Consequently, standard models may systematically misclassify frailty progression in such populations. 

This project proposes a unified framework that addresses both gaps simultaneously. Methodologically, this requires advances beyond standard longitudinal frailty models. While Disease Progression Modelling (DPM) frameworks [4] offer tools for sequencing health events, their assumptions regarding scalar unidirectional progression do not hold for frailty, which exhibits reversibility and multi-dimensional trajectories. There is therefore a gap for a methodological framework that adapts DPM principles to the non-linear, reversible nature of frailty deficits.  By modelling temporal transitions of deficit subtypes, rather than aggregate counts, we can: (1) identify distinct subgroups based on deficit composition and trajectory, (2) generate clinically interpretable strata that could inform future targeted intervention studies, and (3) potentially predict a person’s trajectory and potential adverse outcomes (e.g. cause of mortality, need for specialized care) and intervene earlier. In this way, frailty quantification does not merely optimise outcomes for the majority, but improves detection and intervention across diverse patient trajectories, ensuring equitable care quality regardless of socioeconomic or clinical profile. 

Aims

1. Frailty Subtypes: Establish a clinically meaningful profile of deficit subtypes using a combination of expert input and unsupervised learning techniques, leveraging primary and secondary health records. 

2. Temporal Modelling: Adapt disease progression modelling principles [4] to accommodate the multi-dimensional, reversible nature of frailty trajectories. Evaluate model performance against existing longitudinal FI approaches (e.g. [2]), specifically assessing the ability to predict causespecific adverse outcomes rather than all-cause mortality alone. 

3. Health Inequalities: Apply the developed framework to linked routine health data to quantify how socioeconomic and demographic factors influence deficit composition and trajectory dynamics. Determine whether stratification improves risk detection in relevant subgroups compared to standard aggregate models. 

Training Outcomes

1. Technical/Methodological: 

• Gain proficiency in longitudinal probabilistic modelling and interpretable machine learning for non-linear health trajectories. 
• Acquire advanced skills in routine health data research and understanding within safehaven environments (DataLoch). 

2. Domain Knowledge: 

• Deepen clinical understanding of frailty phenotypes, multimorbidity patterns, geriatric care pathways and patient outcomes. 
• Develop an understanding of the socio-economic drivers of disease progression. 

3. Professional/Leadership Skills: 

• Complete mandatory training in Data Governance, GDPR, and Information Security required for sensitive health data access. 
• Develop capacity for interdisciplinary working, science communication, translating complex methodological findings for clinical, policy, and patient audiences. 

References

[1] Searle, S.D., Mitnitski, A., Gahbauer, E.A. et al. A standard procedure for creating a frailty index. BMC Geriatr 8, 24 (2008). https://doi.org/10.1186/1471-2318-8-24 

[2] Johnson L, Guthrie B, Anand A et al. Accelerated deficit accumulation in frailty and associations with adverse outcomes: a longitudinal population data analysis The Lancet Healthy Longevity, 2026; 7 

[3] L. Johnson, “Understanding Frailty Through the Cumulative Deficit Model: Exploring Dimensionality and Temporality in Routine and Survey Data,” Ph.D. dissertation, University of Edinburgh, Edinburgh, UK, 2025 

[4] Young, A.L., Marinescu, R.V., Oxtoby, N.P. et al. Uncovering the heterogeneity and temporal complexity of neurodegenerative diseases with Subtype and Stage Inference. Nat Commun 9, 4273 (2018). https://doi.org/10.1038/s41467-018-05892-0