Early-detection of high-risk patient profiles using a multistate model

Biostatistics Support and Research Unit

Germans Trias i Pujol Research Institute and Hospital (IGTP)
Badalona, Spain

June 5, 2025

Index

  • Background

  • Objectives

  • Methods

  • Results

  • Limitations & Future Work

  • Conclusions

Background

  • Respiratory infections range from mild to life-threatening conditions and can be transmitted via multiple routes.
  • SARS-CoV-2 and influenza viruses remain leading causes of severe respiratory disease, accounting for a substantial share of global morbidity and mortality.

According to WHO estimates, approximately 116.5 million cases of COVID-19 and 3 to 5 million cases of severe influenza were reported globally in 2023.1, 2

  • Clinical progression in these infections can vary:

    • Some patients remain on general wards; others rapidly deteriorate, requiring ICU care and mechanical ventilation.

Objectives

  • Hypothesis

    • Baseline demographics, presenting clinical features, and early events are key factors in predicting risks of mortality and clinical deterioration.


  • Objectives

    1. Identify clinically relevant prognostic factors associated with mortality in a global cohort of adults infected with respiratory pathogens.

    2. Develop a multistate‐model–based clinical prediction tool to facilitate early identification of high-risk individuals among patients with respiratory infections.

Methods

STRIVE Cohort


  • International prospective multicenter observational study (16 different countries).

  • Adult men and women hospitalized with a respiratory infection (Influenza or COVID-19).

  • Hospitals affiliated with the Strategies and Treatments for Respiratory Infections and Viral Emergencies (STRIVE) research group.

  • Cases ranging from 2013 to 2023.

  • Collected data: demographics, comorbidities, laboratory results and clinical outcomes.

  • Clinical timeline: Start and end dates of different clinical states.

Methods

  • Estimation of event rates: NIV/HFNC, MV/ECMO, O2 support discharge, death and hospital discharge.

  • Exploration of the association between baseline characteristics and mortality.

  • Multistate models:3

  • All analysis were performed using the R software and the mstate package.4

Why multistate models?

\[ h(t \mid X) \;=\; h_{0}(t)\,e^{\beta^{\top} X} \]

  • \(h(t \mid X)\): hazard at time \(t\) given covariates \(X\).

  • \(h_{0}(t)\): baseline hazard.

  • \(\beta^{\top} X = \sum_{i} \beta_{i}x_{i}\): linear predictor.

  • Appropriate when there is a single event of interest.5

\[ h_{A \to B}(t \mid X) \;=\; h_{0}^{\,A \to B}(t)\,e^{\beta_{A \to B}^{\top} X} \]

  • For each pair of states \((A \to B)\), there is a cause-specific hazard \(h_{A \to B}(t \mid X)\).

  • \(h_{0}^{A \to B}(t)\): baseline hazard for the transition.

  • \(\beta_{A \to B}^{\top} X\): covariate effects specific to the transition.

  • Captures multiple possible clinical paths over time.

  • Markov assumption3

Transition probabilities

  • To go from modeling hazard functions to making probabilistic predictions, we need to define:

\[ P_{ij}(s,t) \;=\; \Pr\bigl\{\,X(t)=j \;\big|\; X(s)=i\,\bigr\}, \qquad 0 \;\le s < t. \]

  • \(P_{ij}(s, t)\): probability of transitioning from state \(i\) at time \(s\) to state \(j\) at time \(t\).3

  • Examples:

    • We can predict the probability that a patient is on MV/ECMO two days after admission, given that he was on NIV/HFNC one day after admission.

    • We can also compute the probability of going from Admission to Death within a fixed interval.

Results

Flowchart

Patient profile

Mortality over time

  • Raw mortality rates: Influenza cases = 4.1%    |    COVID-19 cases = 8.2%

Mortality univariate log-binomial models

Mortality univariate log-binomial models

Multistate model

Risk profiles from admission


Risk Profile Chart

Base profile: 60-year-old female of White race with no chronic kidney disease.

Risk profiles: Shiny app 6

The app is available at https://brui.shinyapps.io/STRIVE/


STRIVE Shiny App Screenshot

Markov assumption: NIV/HFNC to Death

Markov assumption: NIV/HFNC to MV/ECMO

Limitations & Future Work

Limitations & Future Work

  • Some transitions (e.g. admission to MV/ECMO) present low event counts, which can lead to imprecise hazard estimates.

  • All results are based on the STRIVE dataset. Internal and external validation are still pending.

  • Markov assumption: violations have been observed in some transitions so we plan to implement a second‐order Markov framework to account for history‐dependent transitions.

Conclusions

  • Multistate models provide a powerful framework for tracking the progression of respiratory infections.

  • Estimating transition probabilities provides clinicians with actionable data to support outcome prediction and more efficient resource allocation for infections such as influenza and SARS-CoV-2.

  • Combining statistical modelling with interactive tools such as Shiny allows for practical, real-time applications in clinical settings.

References

  1. World Health Organization. Coronavirus (COVID-19) Dashboard. Geneva: WHO; 2024.

  2. World Health Organization. Global Influenza Programme: Influenza Update. Geneva: WHO; 2023.

  3. Putter H, Fiocco M, Geskus RB. Tutorial in biostatistics: competing risks and multi-state models. Stat Med. 2007;26(11):2389–2430.

  4. de Wreede LC, Fiocco M, Putter H. mstate: an R package for the analysis of competing risks and multi-state models. J Stat Softw. 2011;38(7):1–30.

  5. Andersen PK, Keiding N. Multi-state models for event history analysis. Stat Methods Med Res. 2002;11(2):91–115.

  6. Chang W, Cheng J, Allaire JJ, Xie Y, McPherson J. shiny: Web Application Framework for R. R package version 1.7.4; 2025.


Thank you!

🎉 Now we can go have lunch! 🎉