gms | German Medical Science

Objectives: Polypharmacy is associated with adverse drug events, which can lead to hospitalization and death. Providing pharmaceutical care can increase drug safety and reduce hospital admissions.

Question to be answered: Can patients receiving polypharmacy who are at risk of hospitalization and in need of pharmaceutical care be identified by using routine data and data mining techniques?

Method: A retrospective database analysis of health insurance data from 2005-2008 and 2007-2010 was performed. Patients aged 18-85 years on continuous polypharmacy receiving 5 or more drugs per quater for 4 quaters were included. Seriously ill patients were excluded. From the first data set (n=44.108) a predictive model was derived from 64 variables in a stepwise approach (CRISP-DM 1.0) using SPSS 19.0 and logistic regression.The final model containing 24 variables was validated using the second data set (n=45.739).

Results: Of 45.739 patients on polypharmacy 39% were admitted to hospital within one year, 88% were on medium polypharmacy using 5-8 drugs per quarter for at least one year. Compared to using the number of drugs as a solely predictor(>13 drugs: n=489, PPV=59.9%) the model identified a larger group of patients with a higher probability of hospitalization and a presumed need of pharmaceutical care (n=1.161, PPV=71.6%). The quality of the predictive model was acceptable (AUC=65.2%, 95% CI 64.7-65.7%) and stable over a two year period. The strongest predictors for hospitalization among patients on polypharmacy appeared to be number of different drugs per year, age, drug costs and the use of metamizol, opioids, loop-diuretics, phenprocoumon und clopidogrel.

Discussion: As cut offs for polypharmacy and the need of pharmaceutical care are inconsistent surrogates had to like hospitalization be used. Therefore the derived predictive model identifies patients at risk rather than risky drugs but still provides hints about which drugs should be monitored more carefully. Routine data are lacking important clinical information to fully use the potential of data mining techniques. Therefore quality of data should be improved to increase quality of prediction, e.g. by extended linkage of primary and secondary data records and intensified documentation of drug use and triggered events by drug experts.

Practical implications: Using data mining techniques help to narrow down mass data and the variables selected deliver a predictive model with interpretable and significant results. The most common definition of polypharmacy using 5 or more drugs as a cut off can be confirmed but is of limited use as the total number of prescribed drugs per year seems to be a better predictor for hospitalization. Still the derived predictive model improves identification of patients on polypharmacy at risk of hospitalization and helps addressing pharmaceutical care more exactly to patients in need.