Abstract
Research based on routinely collected hospital raises several statistical challenges.
Patients are commonly monitored over time to assess disease progression, generating repeated measurements, with potentially complex dependence structures. Moreover, clinical covariates are often strongly correlated, making it difficult to isolate their individual effects by adjustment alone. From a practical perspective, clinicians also need to translate complex multivariable models into interpretable tools and rules for patient profiling, based on a smaller set of relevant covariates and clinically meaningful cut-off values.
This seminar presents two studies conducted at San Raffaele Hospital in Milan, each addressing a different aspect of that problem.
The first study examines five years of telemedicine, asking whether it remained a routine component of care after the pandemic emergency. In this case, dependence has two sources: visits are repeated within patients, and patients are treated across several clinical units. Visit counts were modelled using generalized linear mixed models with a negative binomial response to account for overdispersion. Random intercepts were initially included for patients; subsequently, crossed random effects for patients and clinical units were included in the model. These different modelling approaches help distinguish two phenomena: growth in the number of visits driven by new patients, and increased use among patients already in care. Patient sex and generational cohort were among the factors associated with these patterns.
The second study addresses risk stratification in patients admitted during the COVID-19 surge, with strongly interrelated clinical covariates. Standard regression models were complemented by survival and classification trees to identify data-driven patient profiles and clinically interpretable thresholds. Conditional relationships and interactions among predictors emerge from the tree structure without requiring them to be specified in advance. The tree-based analyses identified three risk strata for in-hospital mortality and three patient profiles for respiratory severity, based on data-driven thresholds involving variables such as age, creatinine and C-reactive protein.
The two studies illustrate different approaches to modelling dependencies among variables. In the first, dependence is explicitly captured through random effects; in the latter, complex relationships among predictors are disentangled through recursive partitioning to identify clinically relevant patient subgroups. In both, dependence is not noise, but part of the structure of the data.