Machine Learning’s Role in Personalized Cancer Prognosis

Using a type of AI known as machine learning, the multicenter research team created a prototype tool called the Cancer Survival Calculator and tested it on a nationwide cancer dataset.

Beyond Cancer Stage

According to Dr. Janczewski, this study aimed to identify the patient, tumor, and treatment characteristics that most greatly influence patient survival for each cancer type.

After cancer experts recommended characteristics to study, the researchers collected relevant information from patients diagnosed in 2015 and 2017 with breast, thyroid, and pancreatic cancers. The patients’ records were part of the National Cancer Database (NCDB), which contains records of 72 percent of newly diagnosed cancer cases in the United States.*

Three-fourths of the collected data were used to train the machine learning algorithms to recognize patterns between characteristics at diagnosis and patients’ survival at five years, and then to rank the factors with the greatest influence on survival.

With the remaining data, the researchers used statistical methods to test the prototype’s accuracy in estimating survival.

Machine Learning and Personalized Cancer Survival

The team included data from 259,485 breast cancer patients, 76,624 thyroid cancer patients, and 84,514 pancreatic cancer patients.

The researchers found multiple characteristics specific to patients, tumors, and treatments for all three cancer sites significantly influenced survival.

The top four factors influencing whether patients were alive five years after diagnosis were as follows, by cancer site:

Breast cancer:

• whether the patient had cancer surgery




• the patient’s age at diagnosis




• tumor size




• time from diagnosis to treatment

Thyroid cancer:

• age at diagnosis




• tumor size




• time to treatment




• lymph node involvement

Pancreatic cancer:

• cancer surgery




• histology, or microscopic analysis of the cancer




• tumor size




• age at diagnosis

Also found to be important for survival from breast cancer were hormone receptor status, which is part of breast cancer staging, and the presence of Ki-67, a biomarker in breast cancer.

Although some of the predictive factors, such as tumor size, are part of cancer staging, Dr. Janczewski said their results showed that many more factors influence survival for cancer patients beyond their disease stage.

Furthermore, their validation testing showed that the calculator was “highly accurate” for all three cancer sites at estimating cancer survival rates – within nine to 10 months of actual survival, Dr. Janczewski reported.

AI’s Personalized Prognostic Power

The Cancer Survival Calculator differs in several ways from cancer survival estimators already in use, according to Dr. Janczewski:

The novel calculator includes specific tumor biomarkers and treatment variables that are known to affect a patient’s estimated prognosis, which Dr. Janczewski said many prior survival calculators lack.

The dataset used to develop the new calculator, the NCDB, is more comprehensive than other calculators’ datasets, according to Dr. Janczewski.

The Cancer Survival Calculator uses new data modeling, such as machine learning, which speeds up processing. The models’ risk predictions also demonstrated improved accuracy compared with predictions generated by older calculators, she said.

AI Tool for Tailored Prognosis

The next steps that Dr. Janczewski identified are to finalize a user interface that will allow the use of the Cancer Survival Calculator in clinical practice, followed by pilot testing the calculator at selected cancer centers.

Eventually, the researchers hope to broaden the calculator by adding all other cancer sites included in the NCDB. Initially, the calculator included breast, thyroid, and pancreatic cancers because these cancers have diverse patient populations and differing frequencies and average survival rates, she said.

The research team plans to make the finalized informational tool available to healthcare practitioners.

Reference :

1. American College of Surgeons Clinical Congress 2023
   – (https://www.facs.org/for-medical-professionals/conferences-and-meetings/clinical-congress-2023/)

Source: Eurekalert