Development and validation of nonattendance predictive models for scheduled adult outpatient appointments in different medical specialties

INTRODUCTION

Nonattendance is a critical problem that affects health care worldwide. Our aim was to build and validate predictive models of nonattendance in all outpatients appointments, general practitioners, and clinical and surgical specialties.

METHODS

A cohort study of adult patients, who had scheduled outpatient appointments for General Practitioners, Clinical and Surgical specialties, was conducted between January 2015 and December 2016, at the Italian Hospital of Buenos Aires. We evaluated potential predictors grouped in baseline patient characteristics, characteristics of the appointment scheduling process, patient history, characteristics of the appointment, and comorbidities. Patients were divided between those who attended their appointments, and those who did not. We generated predictive models for nonattendance for all appointments and the three subgroups.

RESULTS

Of 2,526,549 appointments included, 703,449 were missed (27.8%). The predictive model for all appointments contains 30 variables, with an area under the ROC (AUROC) curve of 0.71, calibration-in-the-large (CITL) of 0.046, and calibration slope of 1.03 in the validation cohort. For General Practitioners the model has 28 variables (AUROC of 0.72, CITL of 0.053, and calibration slope of 1.01). For clinical subspecialties, the model has 23 variables (AUROC of 0.71, CITL of 0.039, and calibration slope of 1), and for surgical specialties, the model has 22 variables (AUROC of 0.70, CITL of 0.023, and calibration slope of 1.01).

CONCLUSION

We build robust predictive models of nonattendance with adequate precision and calibration for each of the subgroups.

Analysis and redesign of a knowledge database for a drug-drug interactions alert system

Physicians tend to ignore drug-drug interactions alerts, this is due to the large amount of irrelevant interactions displayed and the interface in which these alerts are shown. The high rate of clinically inadequate alerts produce "alerts fatigue". This high number of incorrect alerts predisposes physicians to underestimate the electronic prescription systems as useful tools in their practice. We decided to analyze and redesign our drug-drug interactions alerting system knowledge database. In order to do so, we cleaned our knowledge database according to the clinical significance of drug-drug interactions. New drug interactions taxonomy was created in four levels based on clinical significance and the recommendations given in each single monograph of interaction. We proceeded to recategorize the alerts as Active, which present themselves to the physician interrupting the prescribing process, or Passive, which allow physicians to accept the recommendations, and adopt some action in order of minimizing the interaction risks.