Abstract
Background
Antibiotics are the most commonly prescribed drugs for children with estimates that 30%-50% of outpatient antibiotic prescriptions are inappropriate. Most analyses of outpatient antibiotic prescribing practices do not examine patterns within individual clinicians’ prescribing practices. We sought to derive unique phenotypes of outpatient antibiotic prescribing practices using an unsupervised machine learning clustering algorithm.
Methods
We extracted diagnoses and prescribing data on all problem-focused visits with a physician or nurse practitioner between 6/11/2018 – 12/11/2018 for a state-wide association of pediatric practices across Massachusetts. Clinicians with fewer than 100 encounters were excluded. The proportion of encounters resulting in an antibiotic prescription were calculated. Proportions were stratified by diagnoses: otitis media (OM), pharyngitis, pneumonia (PNA), sinusitis, skin & soft tissue infection (SSTI), and urinary tract infection (UTI). We then applied consensus k-means clustering, a form of unsupervised machine learning, across all included clinicians to create clusters (or phenotypes) based on their prescribing rates for these 6 conditions. A scree plot was used to determine the optimal number of clusters.
Results
A total of 431 clinicians at 77 practices with 234,288 problem-focused visits were included (Table 1). Overall, 42,441 visits (18%) resulted in an antibiotic prescription. Individual clinician prescribing proportions ranged from 5% of visits up to 44%. The optimal number of clusters was determined to be four (designated alpha, beta, gamma, delta). Antibiotic prescribing rates were similar for each phenotype across AOM, pharyngitis, and pneumonia but differed substantially for sinusitis, SSTI, and UTI (Figure 1). The beta phenotype had the highest median rates of prescribing across all conditions while the delta phenotype had the lowest median prescribing rates except for UTI.
Table 1. Patient demographics and clinician characteristics
Figure 1. Novel phenotypes of antibiotic prescribing practices across six common conditions
Conclusion
Antibiotic prescribing varies by both condition and individual clinician. Clustering algorithms can be used to derive phenotypic antibiotic prescribing practices. Antimicrobial stewardship efforts may have a higher impact if tailored by antibiotic prescribing phenotype.
Disclosures
All Authors: No reported disclosures
Unsupervised machine learning and pandemics spread: the case of COVID-19
