Modelling and analysis of health care services using regression and Markov models

Abstract

Using data from electronic health records this thesis aims to model and analyse health care services provided to adult patients with chronic conditions. Two aspects of health care services, with unique aims, have been examined.

The first aspect is related to the aim of investigating factors affecting the patients' self experienced quality of the health care encounters with regards to satisfaction, personalized help and general information received. Significant factors were determined by odds ratios resulting from either logistic or multinomial regression, combined with generalized boosted regression. The main findings included: Better self perceived health, increased age, the absence of long lasting illness and not having experienced debasing, accounted positively for the odds of being satisfied. Factors implying a sicker patient increased the odds of receiving help and information; though higher age reduced the odds. Specifically regarding receiving personal help, higher level of education showed an increase in the odds. There were also indications that satisfaction could be negatively correlated with the amount of help and information received.

The second task has been to construct discrete-time patient trajectories, consisting of unique states or events that describe health service usage. Using such patient trajectories this aspect's aim is to model and describe changes and stability in health service usage, and predict future health care events using discrete-time Markov chain and hidden Markov models. Estimation was performed by maximum likelihood and trained by the Baum-Welch algorithm. Both Markov models were justified to describe certain perspectives of health care utilization. Prediction of future health events was only theoretically adequate using hidden Markov models, but its accuracy was unsatisfactory. Also the hidden states of the hidden Markov model, with unknown physical interpretation in a patient trajectory setting, can be induced to represent complex health levels or indices for patients.