Prescription Shopping Anomaly Detection

Wednesday. March 15, 2017 - 3 mins

Overview:

Prescription shopping or doctor shopping can be defined as patients who target prescribers and pharmacies to obtain prescription drugs in excess of medical needs, with a general intent to stockpile, sell, abuse, and/or illegally export. According to the NIH 1 in 148 patients taking an opiate showed suggested patterns of misuse or abuse 1.

I used real world data from the OPTUM Rx DataMart. OPTUM is an integrated database of enrollment, inpatient and outpatient medical claims, pharmaceutical claims, and laboratory results. The data provides information on patient, prescriber, and pharmacy on a fill grain and is de-identified, unlabeled. All duplicates and records with no providers were removed as part of data cleaning.

Method:

From the data and time available the criteria to determine prescription shopping was limited to:

visting too many different physicians within a short time period
switching between providers and/or pharmacies often

Entropy_Formula

I used Shannon’s entropy to determine the unpredictability of a patient’s selection of a provider or pharmacy. In the formula above P_i is the probability of going to a specific provider or pharmacy, calculated by dividng the number of visits to a provider or pharmacy by the total number of visits to providers or pharmacies. Entropy is bounded from 0 to log₂ of the number of total visits to a provider or pharmacy.

A high entropy value lets us know there is a greater chance for a patient to have prescription shopped, and is more suspicious. A low entropy value tells us that is is unlikely that a patient has prescription shopped. Patients with 0 entropy have never went to a different provider or pharmacy.

In order to compare all the entropy values between patients, the provider and pharmacy entropy values were summed (combined entropy).

Unflit_Entropy_Dist

73.8% of patients have a combined entropy of 0. This supports the normal patient behavior, to consistently see the same provider and purchase vicodin from the same pharmacy.

These patients were filtered out to get a more clear look at the ‘abnormal’ patients.

Filt_Entropy_Dist

The distribution of patient entropy shows that patients who are potentially prescription shopping will have a combined entropy greater than 4.

Entropy_Comp

Red = highly suspicious

Orange = moderately suspicious

Green = innocent

I analyized patients in various regions and determined, patients greater than an entropy of 5 are highly likely, patients with an entropy between 4 and 5 are likely to be prescription shopping to be prescription shopping, and patients with an entropy less than 4 are very unlikely to be prescription shopping.

In order to verify the bounds I closely analyized the fills for 120 patients, 40 in each zone to verify the model.

85% in the red zone
50% in the orange zone
2.5% in the green zone

Future Works:

In the future I would like to develop a neural net to determine the bounds for each zone. This would require annotating the data on a patient grain. Furthumore to improve detection I would like to enhance the criteria to look at unusual dosage escalations, inappropriate co-occurrences of items in care, and distance traveled to a provider or pharamacy. Once improving the accuracy of the model, it would be appropriate to extend the algorithm to other opiates.

It also intrests me to look at how legislation in the past 10 to 15 years has affected prescription shopping.

Prescription Shopping Anomaly Detection

Overview:

Method:

Future Works:

Other Topics Learned:

Resources: