4:50 - 5:10 pmSaturday, September 17
Plenary Hall
Personalized medical decision-making
Plenary Hall
Personalized medical decision-making
Stanford University
BackgroundPersonalized medicine, the concept of using patient-specific information to make treatment decisions, has the potential to revolutionize medical care. However, learning which treatments are effective... Read more

Description

Background
Personalized medicine, the concept of using patient-specific information to make treatment decisions, has the potential to revolutionize medical care. However, learning which treatments are effective on which patient subpopulations remains a challenging statistical problem, particularly in the regime of “big data” (availability of many patient covariates).

We develop a new statistical decision-making algorithm for this problem and apply it to the task of warfarin dosing. Warfarin is the most widely used oral anticoagulant agent in the world, but correctly dosing warfarin remains a significant challenge since the appropriate dosage is highly variable among individuals (by up to 10x) due to patient-specific factors. However, an incorrect initial dosage can result in highly adverse consequences such as stroke or internal bleeding. Thus, we tackle the problem of learning and assigning an appropriate initial dosage to patients by leveraging patient-specific factors.

Methods
We formulate the problem of personalizing treatment choices as a multi-armed bandit with high-dimensional covariates. We develop a new statistical decision-making algorithm and prove theoretical guarantees that it achieved near-optimal performance.

Next, we evaluate our algorithm on warfarin dosing using a public patient dataset made available by PharmGKB. The data contains patient-specific optimal warfarin doses for 5,528 patients, as well as 93 patient-specific covariates including clinical factors (diagnosis, medications, etc.), demographics (gender, age, etc.), and genetic information. Our algorithm sequentially assigns low, medium, or high dosages to patients based on their clinical characteristics, observes the patient’s response, and learns the effectiveness of the dosage choice on different patient subpopulations. We compare our algorithm to standard physician practice, which starts all patient on a medium dose (appropriate for most patients), as well as existing bandit algorithms.

Results
Our algorithm assigns the correct dosage for 65% of the 5,528 patients, while the physician policy assigns the correct dosage for only 54% of patients. These gains are primarily achieved because our algorithm correctly doses 57% of low-type patients (33% of the population), while the physician policy always assigns them the (wrong) medium dose.

Furthermore, our algorithm significantly outperforms existing bandit algorithms, particularly when the number of patients is small. For example, when there are 500 patients, our algorithm achieves 60% accuracy while the existing algorithm only achieves 41%. These gains are achieved by our algorithm’s ability to quickly learn in high-dimensional settings (many patient covariates).

Conclusion
Personalized medicine can revolutionize medical treatment decisions. We provide a statistical decision-making algorithm that can aid physicians in making these choices, while simultaneously learning about heterogeneous treatment effects.

Contact Us

We're not around right now. But you can send us an email and we'll get back to you, asap.

Not readable? Change text. captcha txt

Start typing and press Enter to search