Tempus Introduces oFM, a Model That Reads the Treatment Histories of 1.67 Million Patients With Cancer
Tempus Introduces oFM, a Model That Reads the Treatment Histories of 1.67 Million Patients With Cancer
On August 25, the Tempus AI team introduced oFM in the open scientific preprint repository arXiv. The model reads a cancer patient's clinical history as a sequence of dated events. It was developed using a deidentified dataset of 1 672 203 patients. In retrospective comparisons, the patient representations generated by the model were more informative than manually selected clinical and molecular features for predicting risk and ranking patients by the estimated difference in benefit between two treatment strategies.
A medical history consists of events recorded on different dates. A physician takes a biopsy, the DNA test result arrives later, treatment begins, and the tumor's status changes over the following weeks. Each record captures only part of this process. The order of the records traces the course of the disease and its treatment.
oFM converts daily records into a numerical representation of the patient and incorporates digital images of histological sections, which are thin tumor samples viewed under a microscope. RNA data indicate gene activity in the sample. For temporal training, the authors selected the date of a treatment, surgery, or substantial change in the disease, followed by a later date. The model received the patient's history up to the selected day, the conditions of the intervention, and the elapsed time. It then learned to construct a representation of the later state. This approach allows the model to account for the sequence of events in the patient's history.
To separate the value of this representation from the complexity of the algorithm, the researchers paired oFM with a simple linear predictor, a formula that calculates risk from a set of numbers. The same type of predictor received clinical and molecular features selected for each task. For overall survival prediction, the mean AUC, a measure of how well a model distinguishes between patients with different levels of risk, was 0.774 versus 0.563.
In the second test, the authors used 11 groups of patients drawn from previously collected records. Each group compared two treatment strategies. To reduce pretreatment differences between the groups, they applied statistical weighting. Each patient's record received a different weight based on the probability that the patient would be assigned the treatment. The authors then compared which patient representation more effectively ranked people by the expected difference in benefit between the two strategies. oFM performed better in nine of the eleven comparisons.
The authors also examine individual predictions. They remove specific facts from a patient's history and measure how the estimate changes. They then connect influential features in a diagram that reflects when those features appeared. This analysis produces hypotheses about associations that can be tested further in clinical studies.