Last update:
16-Dec-2013

OBJECTIVE In primary care (PC), screening tools are useful for the detection of patients requiring special medical intervention. The objective of this project was to develop a prognostic model for the detection of high risk patients, based on medication history, age, and sex.

METHOD This project was a historical cohort study. All the deaths of the last decade in a small municipality of mountainous Crete were recorded. The age, sex, and most recent medication history were recorded in a data base, for each of these deaths and for most of the living patients in the same municipality. A logistic regression prognostic model was developed, based on the above data, with death as the dependent (binary) variable, and sex, age, and 10 medications or medication groups as the predictors of this model. The 10 medications were selected from among those administered for the most fatal diseases (e.g., coronary artery disease). ROC curves were constructed for all the predictor variables.

RESULTS A total number of 311 deaths were recorded during the ten years in the specific municipality. There were missing data for 14 deaths, so 297 deaths were eventually entered into the data base. The patient histories recorded were 2,045, of which there were missing data for 94, so 1,951 were entered into the data base. The specificity and sensitivity of the model were found to be 97.7% and 81.6%, respectively. The total percentage of correctly classified cases (accuracy) was 96.0%. Significant predictors for the dependent variable, death, were age and use of bronchodilators, sulfonylureas, digitalis, metformin, furosemide, potassium-sparing diuretics, nitrates, calcium channel blockers and ACE inhibitors (p=0.002−0.000). Sex and salicylic acid (as antiplatelet agent) were not significant predictors (p>0.05). The accuracy of the model was finally confirmed using ROC curves (area under curve from 0.537 to 0.673) and the predictor "age" was found to have a cut-off of 77 years.

CONCLUSIONS The model developed here provides general practitioners with a screening test helpful for detecting high risk patients on whom they can focus in order to minimize hospitalization and the risk of death. It is clear that clinicians can develop similar models, using other predictors according to their routine and specialization. A prospective project is currently running to confirm the effectiveness of the model, although time is required for its completion.

Key words: Model, Mortality, Pharmacoepidemiology, Prediction, Primary health care.