Overview I currently work for a large academic hospital as an infectious

Overview

I currently work for a large academic hospital as an infectious diseases (ID) pharmacist and we have recently started an Outpatient Parenteral Antimicrobial Therapy (OPAT) program. The medical director of this program is an ID attending physician with the team comprised of an ID physician, ID pharmacist, and a nurse practitioner. The OPAT team follows all patients discharged on long term parenteral (IV) antibiotics with the overarching goal to improve patient care and limit hospital readmission. Hospital readmissions are a substantial cost for healthcare institutions, largely due to CMS penalties for 30-day readmissions. Patients on home IV therapy represent a higher risk of readmission for many reasons, but largely due to significant infectious risk, drug toxicities, and central line issues. The OPAT program works closely with these patients to solve these problems on an outpatient basis and prevent readmission. In fact, OPAT funding was granted on the basis of cost avoidance. By continuing to show a decrease in hospital readmission rates, the budget for the program continues to be approved by the medical center.

Problem

The OPAT program has continued to grow, with an average census of 80 patients. Given the limited OPAT team members, it’s impossible to preemptively focus care on each patient, but rather resolve problems as they arise. However, the OPAT team does attempt to identify what we would consider a high-risk patient given clinical experience. These patients get additional daily review by the OPAT pharmacist in an effort recognize and resolve any issues that may lead to readmission. While we have significantly reduced readmissions using our current processes, data analytic tools will assist in predicting which patients are at higher risk for readmission and we can further improve our program.

Predictive Modeling

The training data exist for hundreds of past OPAT patients with numerous attributes. The target variable in this training data is whether they were readmitted to the hospital (or ED) at any point in time while receiving IV antibiotics. A classification data mining technique is needed to predict membership to one of two classes. Logistic regression or tree induction can utilize multiple patient specific attributes and construct a model to predict classification (readmission-yes/no). The tree induction model is preferred given its ease of understanding, albeit only if accuracy isn’t significantly reduced compared to a logistic regression model.

Data

The electronic medical record houses accurate and detailed attributes for each patient to be used as training data. These data are available in the data warehouse and can be easily obtained for research purposes.

Attributes to consider:

1. Class of antibiotic (beta lactam, glycopeptide, aminoglycoside, etc.)

2. Renal function (Scr 0-1, 1-2, 2-3, >3)

3. Home health company (internal vs external company)

4. Patient age (can set a threshold)

5. Patient sex

6. Patient race

7. Distance from hospital (> or < 20 miles)

8. Insured (yes/no)

9. Line type (central vs PICC vs none)

Use

The model would be used at the transition of care from discharge to enrollment in the OPAT program. The OPAT team then inserts the attributes into the model to find if the patient meets NO or YES for readmission and a given probability. Once we were able to identify and predict high risk patients, then additional resources could be targeted to those patients. This may include extra labs, more frequent check ins, shorter duration between clinic visits, etc.

Evaluation

The OPAT program has historical readmission data, both prior and since initiation of the OPAT program. Yearly readmission data after the implementation and use of the model could be compared to prior to evaluate the benefits of data science process.

Limitation

This model will be built from training data of previous OPAT patients. Once the model is implemented and high-risk patients are better identified and monitored, new data will not be able to be continuously used to update the previous model – as the method of high-risk patient identification has changed.