AI selfie tool predicts cancer survival and raises privacy concerns

Researchers have developed an AI tool that can predict a cancer patient's chances of survival from a single selfie. Dr Raymond Mak, a radiation oncologist at healthcare research company Mass General Brigham, said the system uses algorithms to analyse a simple face photograph and measure more than one health signal, including biological age and mortality risk.

The team said the two algorithms look at different parts of the face. The age algorithm focuses on the temples, the forehead and the nasolabial fold, the line from the nose to the corner of the mouth. The mortality risk algorithm, called face survival, looks at the bridge of the nose and the area under the eyes. Dr Mak described the system as a little bit of a black box, but said techniques allowed the team to ask the algorithms which parts of the face they used. When combined for cancer patient groups, the two measures complemented each other and added prognostic information beyond what is usually used in the clinic, he said.

The age prediction algorithm was trained on 700,000 photographs where the person's age was known. The mortality algorithm was trained on more than 30,000 photographs of cancer patients where researchers knew how long they lived and whether they died. The initial algorithms were trained on 40 million face photographs.

Dr Mak said the work could mark the start of a new field, which the team calls facial health recognition AI. He said the platform could be trained to predict other conditions or outcomes that may be reflected in the face, such as Alzheimer's disease, dementia, or a person's blood count. The team is already training an algorithm to predict blood count from a simple face photograph.

The research also raises privacy concerns. Dr Mak said the lab debated whether to do the work because photographs are everywhere and AI can analyse anything on the internet. The team decided to move forward to make people aware that medical insights can be gleaned from such photographs, and to ensure regulators and ethicists know what AI can do so that guardrails can be put around safe and medically helpful use.