gms | German Medical Science

Background: The situation in health care and health research is changing. Nowadays, there is increasing interest on research results which are related to daily routine in medical care. In order to put clinical trials into a more ordinary setting, approaches should be forced which enable the participating patients to flexibly contact the trial physicians for treatment and process monitoring. The available tools of information and communication technology offer several options to realize e.g. a web-based contact or a response system between patients and physicians, which offer highly flexible designs for process monitoring programs. From a methodological point of view this kind of patient data requires suitable analysis methods for several reasons.

Research question: The fact that repeated measurement data arise from such response or progress monitoring systems entails the challenge to properly handle measurements from varying intervals, especially intervals which are not equidistant. Moreover, one usually don’t seek to investigate variable characteristics at different time points when analyzing such kind of data, but rather to find patterns of trajectories which display the development of interesting outcome variables over time and may be clustered in several generic classes. The main objective of this proposal is therefore to present non-standard statistical methods available for the analysis of longitudinal data arising from response monitoring systems or large secondary data sets.

Methods: The three methodological approaches of sequence pattern analysis (SPA), hidden markov models (HMM), and growth mixture models (GMM) will be presented and discussed critically. A focus is set on their presumptions and limitations in practice. Moreover, a simulation study based on a large secondary data set in the field of rehabilitation medicine will be performed. In this study monthly counts of medical visits shall be simulated for two health states: “need for rehabilitation” and “no need for rehabilitation”. The assumption is made that patients with “need for rehabilitation” use the health care system more often than the others. To consider a realistic situation, additional error counts which arise from visits because of other diseases are included. The counts shall be modelled via Poisson distributed random variables.

Results: The HMM and GMM approaches can use the count data directly. For SPA the counts are summarized in categories (by five counts). The simulation study considers two situations: patients either stay in one state over the whole time period, or patients change from “no need for rehabilitation” to “need for rehabilitation” in random time intervals. Finally, the proportion of correctly classified patients (at the end of the time period) is compared between the approaches.

Discussion: The simulation study shall give a clue how the approaches react in the different situations. In the real data set more information about the visits – prescriptions, treatments, and diagnoses – is available, which is why further simulation studies should be performed to evaluate the best performing approach for more complex situations using more or all information in the data set. Moreover, modifications of the intensities of case and error counts and the impact on the classification should be examined as well. Overall, health services research will benefit with respect to a more precise estimation of interesting parameters in cases where treatment pathways are to be analyzed.

Practical implications: In context of rehabilitation medicine, treatment patterns which may indicate demand for rehabilitation will help the health care insurance to identify patients with respective needs more easily. These patients could particularly be contacted, and rehabilitation could be initiated timely. This would improve both allocation of resources in health care as well as associated costs, and the outcome of patients.

Contributed equally: B. Mayer, J. Dannenmaier