When a person is diagnosed with cancer, it is important to start treatment as soon as possible. But the problem that oncologists face here is determining which is the best treatment that can be prescribed, since, although it is thought that one can always go for the most aggressive drug, the reality is that it can be fatal for the patient.
the state A patient’s general condition is one of the most critical factors when a team decides on treatment, since talking about a very old and frail person is not the same as talking about a young and trained boy. Here the assessment depends on the expert’s clinical eye, but now they are working on an AI to be able to read our faces with amazing accuracy to draw a profile.
The system is named FAHR-Face and a priori attempts to estimate mortality risk and biological age of cancer patients based solely on facial photographs.
Work. To get a machine to understand the ravages of disease in the human face, scientists didn’t start from scratch. Here the base model was trained using over 40 million facial photographs and from here the AI was split to have two models trained on 34,363 photographs of oncology patients:
- FAHR-FaceAge: Designed to calculate biological age (how old the body actually appears after date of birth).
- FAHR-FaceSurvival: Specially trained to predict overall mortality risk.
Review. To verify that the system really worked and wasn’t just memorizing data from his home hospital, the model underwent independent validation on a different cohort with nearly 5,000 patients. Additionally, the researchers confirmed that the AI was robust and would not be fooled by variables such as makeup, lighting, poses or even cosmetic surgery.
How he does it. According to research, to estimate biological age, artificial intelligence focuses its “attention” on nasolabial folds, forehead and temples. However, when it comes to predicting survival, the algorithm looks at the lower part of the eyes and the bridge of the nose.
Something that works pretty well, since patients in the highest risk quartile classified by AI have a 3.23 times higher risk of death than those in the lowest risk quartile. The 122 patients classified as high risk translated into absolute figures, 52.5% died within 90 days, showing that this was quite close to reality.
It is a support. Although we are talking about important statistics, the reality is that experts ask to be careful, since this system does not diagnose cancer or give patients an exact “expiration date”. These are purely statistical estimates applied to individuals who already have an oncological diagnosis.
The reality is that the technology still has important limitations to overcome. The current data comes from previous studies based on cohorts where the patients were mostly white, which requires continuing to train the model so that it does not discriminate based on skin color. Furthermore, detailed information on treatment-specific toxicities is lacking and, most importantly, it has not yet been proven through prospective clinical trials that the use of this AI improves treatment decisions or prolongs the actual survival of patients.
Picture | Cottonbrough Studio
In Xataka | Health will prioritize young cancer patients to use the machines donated by Amancio Ortega
