Surgical Risk Calculator: POTTER
A highly accurate and understandable risk predictor trusted by top surgeons
Surgical risk is not linear
Emergency surgery, which deals with acute threats to life, accounts for roughly 7% of surgeries in the United States. It is critical for surgeons to be able to reliably estimate the risk of mortality and complications prior to committing to surgery.
Given the wealth of data available, data-driven risk estimates can augment a surgeon's intuition. The existing tool is a calculator that requires the surgeon to enter 20+ variables and returns a single predicted risk value with no additional context. This calculator is based on logistic regression and as a result the patient's risk factors are used to make predictions linearly and independently, without any interaction between different risk factors.
Given the wealth of data available, data-driven risk estimates can augment a surgeon's intuition. The existing tool is a calculator that requires the surgeon to enter 20+ variables and returns a single predicted risk value with no additional context. This calculator is based on logistic regression and as a result the patient's risk factors are used to make predictions linearly and independently, without any interaction between different risk factors.
NSQIP Surgical Risk Calculator based on logistic regression
Easy to use and understand
We collaborated with Massachusetts General Hospital (MGH) trauma surgeons to build a surgical risk calculator using interpretable methods. It is used for training and daily rounds at MGH and other top hospitals in the world, with over 500 users. The calculator only asks relevant questions which help with doctors’ understanding. In this case, the questions are often less than 8 compared to over 20 previously. The logic to arrive at the estimates is clear and can be used for counseling patients.
Validated by domain experts
Because the model is a simple decision tree, doctors who are not machine learning experts can understand and validate the medical meaningfulness of each branch and prediction. This level of transparency establishes trusts in the medical community where blackbox models cannot.
Our analysis was presented at the American Surgical Association's 138th Annual Meeting and published in Annals of Surgery, a top surgical journal.
Our analysis was presented at the American Surgical Association's 138th Annual Meeting and published in Annals of Surgery, a top surgical journal.
Part of the Optimal Decision Tree predicting mortality after surgery.
Used and tested across the world
The tool has been proven useful for the following purposes, some of which are being measured in clinical trials around the world:
- Improved quality in physician decision making
- Better patient counseling
- Training tool for residents and junior doctors
Unique Advantage
Why is the Interpretable AI solution unique?
-
Ability to capture clinically-relevant interactions between variables
The superior modeling power of Optimal Trees allows capturing non-linear effects without sacrificing interpretability
-
Simplicity and interpretability of calculator
A simple user interface is backed up by transparent evidence that doctors can trust and use with ease