ESMO 2021: AI, Targeted Data and Predictive Analytics to Monitor and Proactively Manage Patient Outcomes and Validate the Quality and Benefit of Cancer Care

( In this presentation, Dr. Mozzi Etemadi discussed how computer programs, often referred to as artificial intelligence, can be used for cancer-based applications such as cancer screening.

He began by using the example of training a computer program to recognize the suit on playing cards. The question posed is: “Is this suit A or B”. Based on which choice the computer chooses, and whether this is right or wrong, the computer modifies its representation of suit. Generally, representation is defined as how the world is viewed as shaped by information that is correct or incorrect. Computer representations can be quite powerful if provided enough training information but are limited by human factors such as the dataset and design of the test, and so are not good at predicting things they have not seen before. However, with enough information, computer representations can outperform humans at tasks such as cancer diagnosis.

In contrast, humans use broader representations than computer programs to understand and interact with the world. Humans can represent things using context gleaned from various aspects of the world such as other humans, life experience, education, and language. For example, human radiologists have varied representations they use to diagnose cancer on a CT scan. Based on their training and experience, they can integrate observations such as nodule size, shape, and other characteristics, to classify a nodule as likely malignant or benign.

Computers can be trained with a program, termed artificial intelligence, to perform cancer screening. This takes large training sets to provide enough information for these programs to perform accurately. Dr. Etemadi provided two examples to prove this point. By training a computer using 42,290 examples of lung cancer on CT scans of the chest, an artificial intelligence program was able to diagnose lung cancer with 11% fewer false positives and 5% fewer false negatives. Similarly, by training an artificial intelligence program on 121,455 mammograms, the program was able to diagnose breast cancer with 5.7% fewer false positives and 9.5% fewer false negatives.

Such successful efforts raise many caveats and limitations. These programs can only tell you what they are programmed to tell you, in this case, whether cancer is likely present or not. They cannot provide more information unless they are trained to do so, and training such programs requires a very large amount of input data. Additionally, the optimal method of handling the results of these programs within clinical workflows remains to be determined. Should they prioritize findings for subsequent human review? Should they automatically trigger a biopsy? Finally, broad prospective validation will be necessary before these are regularly incorporated into clinical decision-making.

Presented by: Mozzi Etemadi, MD, PhD, Northwestern University

Written by: Alok Tewari, MD, PhD – Genitourinary Medical Oncologist, Instructor in Medicine, Dana-Farber Cancer Institute, Harvard Medical School, Twitter: @aloktewar during the 2021 European Society for Medical Oncology (ESMO) Annual Congress 2021, Thursday, Sep 16, 2021 – Tuesday, Sep 21, 2021.

email news signup