AI can predict survival in cases of colorectal cancer – 11/29/2023 – Science

A study published in the journal Scientific Reports shows that models based on artificial intelligence (AI) are capable of predicting the death and survival rates of patients with colorectal cancer with almost 80% accuracy. According to the authors, the results indicate that these tools can be useful for planning and evaluating health services, as well as guiding referral protocols.

Colorectal tumors are one of the most common types of cancer in the world, with almost 2 million new cases per year and an increasing trend in this number. In Brazil, data from the National Cancer Institute (Inca) indicate the emergence of 44 thousand new cases annually, 70% in the Southeast and South regions.

Machine learning techniques based on artificial intelligence have been increasingly used to predict mortality and survival rates due to their ability to improve automatically, without the need for constant programming — unlike statistical models used in epidemiological studies which, as reality changes, they become obsolete and less accurate.

The study, supported by Fapesp within the scope of the project “Cancer Control in the State of São Paulo (ConeCta-SP): from knowledge to action”, which designs strategies to control the disease in a short space of time, was one of the first to be carried out predicting the survival of cancer patients based on a large database using AI and verifying the validity of these models in Brazil. The work involved groups from the Oncocentro Foundation of the State of São Paulo (Fosp), the Faculty of Public Health of the University of São Paulo, the AC Camargo Hospital and the Mauá Institute of Technology.

De-identified information on socioeconomic status, clinical and care characteristics and survival of 31,916 colorectal cancer patients treated in more than 70 hospitals in the state of São Paulo between 2000 and 2021 belong to the Hospital Cancer Registry of the State of São Paulo, managed by Fosp.

The researchers evaluated and compared the prediction validity of three AI algorithms: Random Forest, Naive Bayes and XGBoost. The latter presented the best result, correctly predicting 77% of deaths and 77% of survival (at 1, 3 and 5 years from tumor diagnosis).

“The performance of the three models showed a hit rate between 76% and 77%”, says Lucas Buk Cardoso, researcher at the Embedded Electronic Systems Center (NSEE) at the Mauá Institute of Technology and first author of the study. “In addition, it was possible to obtain data from the most important patients for predictions — placed as inputs for AI —, allowing a better understanding of the size of the impact of this information and validation with the knowledge already disseminated in the area.”

The most important input in all models was the clinical staging of the cancer, which contains information about the degree of the disease — the more advanced, the more decisive for predicting death. Other important information was related to the treatment carried out, such as surgery, radiotherapy and chemotherapy, in addition to the presence of recurrence, indicating whether the cancer returned or not, the patient’s age and year of diagnosis of the disease.

The variables that best predicted survival in XGBoost were clinical stage, surgery performed, hospital treatment, age and year of diagnosis.

Practical and methodological advances

According to the authors, the study has the potential to be the first of many that will allow the simulation of scenarios and impacts on the survival of cancer patients. With the information obtained, better clinical and management decisions in public health can be made.

“This type of assessment can indicate models that will serve as instruments for decision-making by managers in moments of potential disruption in health services, as happens in pandemics, for example”, explains Tatiana Toporcov, professor at FSP-USP and co-author of the article.

Vanderlei Cunha Parro, professor at the Mauá Institute of Technology, also highlights the nature of evaluating the boundaries between statistical methods and those that use machine learning. “Such an investigation may also give rise to a methodological review, with new data to be included and others to be excluded.”