Heart Failure Trajectory Markov Chain Modelling and Cost Study

PASS information

Title: Heart Failure Trajectory Modelling and Cost Study

Protocol version identifier: 1.0

Date of last version of protocol: 22.02.2023

Country(-ies) of study: Estonia, TBA

Authors: Markus Haug, MSc; Raivo Kolde, PhD

2. LIST OF ABBREVIATIONS

Abbreviation	Meaning
CDM	Common data model
HF	Heart failure
OHDSI	Observational Health Data Sciences and Informatics
OMOP	Observational Medical Outcomes Partnership
SoC	Standard of care
TM	Telemonitoring
UTARTU	University of Tartu

3. RESPOSIBLE PARTIES

3.1. INVESTIGATORS AND AUTHORS

Investigator/Author	Institution/Affiliation
Markus Haug	University of Tartu
Raivo Kolde	University of Tartu

4. Abstract

Real-world health data (RWD) has the potential to transform health economic research providing more accurate and representative models of treatment practice and outcomes. To facilitate the adoption of RWD for health economic modelling, we have created two R-packages to learn and evaluate Markov models on OMOP CDM formatted data sources. This will allow to learn such model in easily distributable, transparent and reproducible manner. To demonstrate the utility of the approach we are conducting a federated network study to replicate a published Markov model for the progression of heart failure. The goals of the study are to explore the variability of the resulting models and associated costs, the effect these differences have for the economic analysis and the ways how we can integrate the information from multiple sources.

5. AMENDMENTS AND UPDATES

Number	Date	Section of study protocol	Amendment or update	Reason
V1	02.2023	All	Launch	Initiation

6. RATIONALE AND BACKROUND

A range of modeling techniques have been used in health economics research, but cohort-based Markov models have been the most commonly used methods in health technology assessment, as they are relatively simple to develop, debug, analyze, and communicate¹. Markov models are described in terms of the conditions that individuals can be in (“health states”), how they can move between such states (“transitions”), and how likely such moves are (“transition probabilities”) within a given time period (“cycle length”).

The models are typically created based on published information on event frequencies, but sometimes information from real world cohorts is used as well. To get more appropriate approximation of the situation in particular locations, real-world data should be used much more. Data sources in OMOP CDM offer a great opportunity to learn the models in this manner. Therefore we developed R packages for developing Markov models on OMOP formatted datasets. The packages build a layer on top of the current OMOP infrastructure of building cohorts (as “health states”) and connecting to the OMOP database to learn such models. We have also created a Shiny app to pool together the results from multiple data sources and enable comparison of those.

The study we replicate was published by Thokala et al in 2020² on heart failure. This paper modelled patients’ observational health data with Markov Chains and used cost-effectiveness analysis to determine whether Telemonitoring (TM) would be a feasible new treatment in comparison of the standard of care (SoC). In the paper they were contingent on transparent methods for making important medical decisions (whether or not start using TM in treatment). The goal of the original study was to demonstrate that real world datasets can provide evidence for economic modelling. In this study we go one step further showing that we can learn the models in a network of real world data sources, while being transparent and reproducible in methods, phenotype definitions and results. This study would provide valuable insights that would aid in the ongoing development of the packages.

Heart failure (HF) currently accounts for 1% to 2% of the annual healthcare budget in most developed countries and is associated with high levels of morbidity and mortality. TM can facilitate early detection of clinically significant changes and earlier intervention to re-stabilize the syndrome and prevent emergency admissions. There have been studies of cost-effectiveness of TM compared with usual care for HF, estimated using modeling, to help decision makers assess value for money. Here we want to see how much the model conclusions change depending on where the model parameters are learned.

7. RESEARCH QUESTIONS AND OBJECTIVES

The aim of this study is to model the comparative heart failure patients’ movement between the disease acuteness states via health trajectories built by R package Cohort2Trajectory. The Markov Chain modelling and monetary analysis of the states will be conducted by the R package TrajectoryMarkovAnalysis.

Specifically, the study has the following objectives:

1. Comparing and assessing heart failure Markov chains and treatment costs between data partners.

2. Assessing whether or not telemonitoring would be a feasible alternative treatment method to the standard of care.

3. Introduce the method packages Cohort2Trajectory and TrajectoryMarkovAnalysis to the OHDSI community.

8. REASEARCH METHODS

The study will be conducted on OMOP CDM version 5.3. We will define the target cohort, HF state cohorts and learn Markov model parameters as well as state cost parameters. Later we will use TM hazard ratios from earlier publications to calculate corresponding Markov Chains and then later conduct cost-effectiveness analysis.

8.1 STUDY DESIGN

We start with constructing the target cohort and after that the health states HF0, HF1, HF2, HF3 and HFD as defined before. Using the package Cohort2Trajectory we will construct trajectories for each patient showing their status monthly. The observation period of each patient is 5 years before the subject’s death. That is 60 months, meaning that each patient will contribute to the Markov chain for 60 cycles (see Figure 1).

The Markov chain’s parameters (transition probabilities) will be calculated using the maximum likelihood estimation. This and the states’ cost analysis will be conducted by the TrajectoryMarkovAnalysis package. The states costs will be queried from the OMOP CDM cost table. If the data partner has not populated the cost table they can still participate in the study contributing with the Markov chain parameters. Individually the packages output linear patient treatment trajectories and Markov chains (see Figure 2) with cost statistics respectively.

Figure 2: Example of Markov chains

8.2 SETTING

All of the patients have been divided into monthly states which are constructed as follows:

1. HF0 - the patient has had zero hospitalizations in relation to heart failure the past year;

2. HF1 - the patient has had one hospitalization in relation to heart failure the past year;

3. HF2 - the patient has had two hospitalizations in relation to heart failure the past year;

4. HF3 - the patient has had at least three hospitalizations in relation to heart failure the past year;

5. HFD - the patient died during the ongoing month;

8.2.1 STUDY POPULATION: TARGET COHORT

All patients included in the study:

1. are 18 years or older;

2. have had at least 1 hospitalization in relation to heart failure;

3. die at least in five years after having first heart failure diagnosis (observation period 5 years);

Cohort entry event:

Earliest visit occurrence five years before death (see Figure 3).

Cohort inclusion criteria:

1. Having at least one death occurrence;
2. Having at least one impatient visit related to a diagnosis of heart failure;
3. Patient is at least 18 years old before being included in the target cohort.

Cohort exit criteria:

Patients exit the cohort at the time of death (five years after inclusion).

Figure 3: Observation timeline

8.2.1 STUDY POPULATION: STATE COHORTS

STATE COHORT #1: HF0 Cohort entry event:

1. Any visit having zero hospitalizations in relation to heart failure in the last 365 days.

Cohort inclusion criteria:

1. Having at least one death occurrence;

2. Having at least one impatient visit related to a diagnosis of heart failure;

3. Being at least 18 years old before cohort inclusion.

Cohort exit criteria:

1. End of continuous observation or any death or any occurrence of visit related to heart failure.

STATE COHORT #2: HF1

Cohort entry event:

1. Any visit having one hospitalizations in relation to heart failure in the last 365 days.

Cohort inclusion criteria:

1. Having at least one death occurrence;

2. Having at least one impatient visit related to a diagnosis of heart failure;

3. Being at least 18 years old before cohort inclusion.

Cohort exit criteria:

1. End of continuous observation or any death or any occurrence of visit related to heart failure.

STATE COHORT #3: HF2

Cohort entry event:

1. Any visit having two hospitalizations in relation to heart failure in the last 365 days.

Cohort inclusion criteria:

1. Having at least one death occurrence;

2. Having at least one impatient visit related to a diagnosis of heart failure;

3. Being at least 18 years old before cohort inclusion.

Cohort exit criteria:

1. End of continuous observation or any death or any occurrence of visit related to heart failure.

STATE COHORT #4: HF3

Cohort entry event:

1. Any visit having three or more hospitalizations in relation to heart failure in the last 365 days.

Cohort inclusion criteria:

1. Having at least one death occurrence;

2. Having at least one impatient visit related to a diagnosis of heart failure;

3. Being at least 18 years old before cohort inclusion.

Cohort exit criteria:

1. End of continuous observation or any death or any occurrence of visit related to heart failure.

STATE COHORT #5: HFD

Cohort entry event:

1. Any death having at least one hospitalization in relation to heart failure five years before dying.

8.3 DATA MANAGEMENT & DATA ANALYSIS

The package relies on the OMOP CDM. For running the study patient level data on heart failure diagnosis and death is required. Data on costs is not necessary but recommended for further analysis. After running the study summarized tables will be created into results folder, please share ONLY the results folder.

The described patient cohorts will be generated automatically during the execution of the study. Tables needed for the study will be created into specified schema and dropped after a successful execution of the package workflow. The analysis of the cohort data will be done by R-packages Cohort2Trajectory and TrajectoryMarkovAnalysis using a predefined study. The output report consists of description of the database, Markov chain matrix of defined HF states, matrix of log-rank tests between transitions (observed data vs generated data), summarized demographics table, summarized sunburst plots of patients’ trajectories, state cost statistics and cost-effectiveness reports (ICER for SoC vs TM).

9. PROTECTION OF HUMAN SUBJECTS

The study uses only de-identified data. Confidentiality of patient records will be maintained at all times. Data custodians will remain in full control of executing the analysis and packaging results. There will be no transmission of patient-level data at any time during these analyses. Only aggregate statistics will be captured. Study packages will contain minimum cell count parameters to obscure any cells which fall below allowable reportable limits. All study reports will contain aggregate data only and will not identify individual patients or physicians.

10. MANAGEMENT AND REPORTING OF ADVERSE EVENTS/ADVERSE REACTIONS

This study will provide a descriptive summary of individuals having hospitalizations caused by heart failure and dying within five years of observation period. For each database a Markov Chain and aggregated cost analysis will be reported as well as metrics of the learned Markov Chain performance.

11. PLANS FOR DISSEMINATING AND COMMUNICATING STUDY RESULTS

Dissemination activities will be of a scientific nature (articles in scientific journals, presentations at conferences, etc.). Our aim is for these studies to be made available as soon as possible in order to support treatment decisions in the global medical arena.

12. REFERENCES

1. Thokala, P., Dodd, P., Baalbaki, H., Brennan, A., Dixon, S. and Lowrie, K. (2020). Developing Markov Models From Real-World Data: A Case Study of Heart Failure Modeling Using Administrative Data. Value in Health, 23(6):743–750. https://doi.org/10.1016/j.jval.2020.02.012.

2. Caro, J.J., Briggs, A.H., Siebert, U., et al. (2012) Modeling good research practices—overview: a report of the ISPOR-SMDM Modeling Good Research Practices Task Force–1. Value Health, 15(6):796-803. https://doi.org/10.1016/j.jval.2012.06.012.

---

1. Modeling good research practices—overview: a report of the ISPOR-SMDM Modeling Good Research Practices Task Force–1↩︎

2. Developing Markov Models From Real-World Data: A Case Study of Heart Failure Modeling Using Administrative Data↩︎