Mammograms are a common and routine test used to screen for breast cancer.
When a breast tissue biopsy is recommended following a suspicious finding on mammogram, Amino’s data indicates that about a third of all biopsies occur on the same day. As a patient, you may prefer a doctor who has a higher rate of same-day biopsy following mammogram, because early diagnosis of conditions detected by a biopsy is important for effective treatment.
To help you understand how your doctor’s rate of ordering a biopsy the same day as a mammogram compares to a typical doctor who performs mammograms:
- We first measure the actual same-day biopsy rate for your doctor’s patients whose mammograms included a suspicious finding that merited a biopsy.
- We then analyze the typical practices of all other doctors in the United States who perform mammograms, using a data-driven model to predict the rate of same-day biopsy for patients like the ones who see your doctor.
By comparing these two numbers, you can see whether your doctor’s rate of same-day biopsy is lower than, similar to, or higher than predicted.
This prediction adjusts based on the types of patients the doctor sees, so a doctor who treats sicker patients may have a different predicted rate than a doctor who treats healthier patients.
Methodology Deep Dive
Amino compares the predicted rate of a procedure to the actual rate measured for a specific medical professional. Our predictive model incorporates data about every patient in our database who received the procedure, including the patients’ demographic, diagnostic, and place of care data. Therefore, we are able to calculate a risk-adjusted decision factor analysis on a per-provider basis.
To calculate this analysis for an individual doctor, we first pass data about that doctor’s patients to a model for prediction. The model produces an overall prediction of the decision factor rate we would expect to see for these patients if they went to a typical doctor in the US. The model that we use is a multivariate logistic regression, where the patient and facility information are features in the model used to predict the decision factor of interest.
It is important to note that we do not compare an individual doctor’s rate to the overall average rate for all doctors, as this would fail to account for the types of patients the doctor sees and how sick those patients are. Instead, we pass information like age, sex, and diagnosis to the statistical model so that these risk factors may be adjusted for in the prediction. This means that doctors who have higher-risk or sicker patients are not unfairly penalized in our analysis.
For each doctor about whom we report a decision factor, we compare our predicted rate to the actual rate measured and classify the doctor’s rate as lower than, similar to, or higher than predicted. Specifically for mammogram same-day biopsy rates, we report that a doctor’s same-day biopsy rate is lower or higher than their predicted rate if we are 95% confident that the doctor’s rate has at least 5% difference from the rate predicted by the model.
We determine confidence using a beta distribution for the actual rate of the procedure performed by the doctor and then testing if the cumulative probability of the mean predicted probability, when the desired difference interval is included, meets the 95% confidence threshold.
In addition, we do not report a decision factors analysis for a doctor if we do not have enough data to be confident in our assessment. To determine the threshold for the number of data points we need to provide a decision factor analysis for a doctor, we take the overall rate of the decision factor of interest, take into account the desired difference interval, and then add one event at a time to the alpha shape parameter where the beta shape parameter is one, until the confidence threshold is reached. This process is then repeated, adding the counts to the beta shape parameter instead. The more conservative of the two numbers is used as the minimum number of data points required to report on the decision factor.
The data we use to build our logistic regression model is determined by the health condition or medical procedure we are analyzing, and we use clinical expertise to identify the specific procedure codes relevant to what we are measuring.
In all our data analyses, Amino strives to responsibly present facts that help everyone make decisions about their care.