Description

An analysis model for the MBDS allows for the detection of interesting information in order to study treated case mix, its behaviour, and the different healthcare profiles that appear in the hospitalisation process. This model has several essential components that facilitate the knowledge of the case mix of the hospitalisation records present in the national MBDS:

• Dimensions: General objectives of knowledge.
• Analysis axis: Dimensional analysis lines.
• Indicators: Processed data that inform about concrete facts of one or more dimensions.
• Classification levels: Recommended terms for indicator elaboration, in a way that allows aggregation and disaggregation.
• Filters: Selection conditions that restrict indicator calculus to a sample of the recorded universe. Selection can involve one or more variables.

There are seven dimensions for analysis:

• Descriptive: frequency distribution for incidence analysis that facilitate the description of the series.
• Attendance: usage rate for spontaneous demand or suggested by a medical professional of the population of health centre admissions.
• Average resoluteness: assistance duration and contact number.
• Clinical practice styles through studies that assess the variability level.
• Patient security: bad practice or avoidable through prevention effects.
• Clinical effectiveness: health restoring without side effects.
• Efficiency: Performance with resources in health assistance.

There are also six analysis axis:

• Point cut: Descriptive.
• Temporal series: Descriptive of one year with relative differences with a three year period.
• Base value: Comparatives with a year selected as the base value
• Standard intra-series: Comparatives with the mean value of the series' totals.
• Best practices: Using hospital groups, compares values from the whole dataset with values below the 25-percentile of the series.
• Adjustment lines: Recommended as complementary analysis axes according to previously identified variables for each indicator.

Web App

One of the project outcomes was a website in which the healthcare quality indicators can be interactively consulted using the described analysis axes. The application analyses over 40 million records and several tens of indicators, performing calculations in real time.

A new methodology of indicator adjustment based on Bayesian Networks stands out, developed by the Data Mining Group of the University of Cantabria:

Risk analysis methodology (sometimes named severity or case-mix adjustment) is a fundamental part in healthcare quality assessment and is used routinely in several national and international projects of service providers quality (ranking, budget assignment, etc.). For example, the AHRQ publishes its quality indicators along with generic risk adjustment results. However you can't avoid all the variability and further analysis is needed for a practical concrete problem, like budget assignment.

The most used adjustment methodology found in literature is based on logistic regressions with the fundamental limitation of only being able to reproduce monotonous relationships between the adjustment factors and the adjusted indicator. The proposed methodology based on Bayesian Networks overcomes this limitation being the ideal technique for modelization indicators depending on discrete variables:

When factors are discrete, Bayesian Networks are the most efficient solution for building probabilistic models of all the variables, allowing to estimate the probability of the indicator from the known risk factors: P(yk ,x1k ,xnk). In this case, the Bayesian Network offers an alternative more efficient for adjusting the indicator using the different profiles and its probabilities.

Our contribution

Predictia has participated in several phases of the project. Our contribution has been focused on:

• Evolutionary development of the indicator analysis application: an application developed in Java using frameworks like Struts, Spring e Ibatis between others.
• Development of a data flow system for new data ingestion and indicator calculation: an ETL system developed in Java capable of managing the high data volumes involved in the project.
• Evolutionary development of the statistical indicator adjustment model model based on Bayesian Networks.