THE‘artificial intelligence promises faster, more precise, personalized diagnoses. But what happens if algorithms, like human beings, end up being influenced by invisible prejudices? According to a recent study led by Kun-Hsing Yufrom the Harvard Medical School and of theBlavatnik Institutethe AI used for cancer diagnosis can make errors related to demographic discrimination. This was reported in a press release HANDLE of December 16th. A result that surprises – and worries – because it concerns one of the most delicate areas of contemporary medicine:analysis of tumor samples.
When the diagnosis is not so “objective”
Histopathological analysis of tumors (the microscopic analysis of a tissue sample taken from the body via biopsy or surgery) is considered a highly objective evaluation: by looking at cellular images, a doctor cannot infer anything about the patient’s age, sex or ethnic origin. Yet, according to the study cited by ANSA, AI instead manages to obtain precisely this informationusing them – unconsciously – in the diagnostic process.
Analyzing the results of four of the most used artificial intelligence modelsresearchers found that the algorithms can infer demographic data from samplesgenerating distortions in diagnosis. An unexpected effect: «We would expect the pathological evaluation to be objective», explains Yu. “When evaluating images, it is not necessary to know the patient’s demographics to make a diagnosis.”
Mistakes: who risks more
The problem is not theoretical. The biases (errors, prejudices) identified have concrete consequences on the quality of diagnoses. As ANSA reports, some models show greater difficulty distinguishing lung cancer subtypes in African American and male patientsor in correctly recognizing breast cancer subtypes in younger patients. In other words, AI can be less accurate in some specific groupsamplifying already existing inequalities in access to care and quality of healthcare.
Where do biases come from?
According to the authors of the study, the origin of these distortions is to be found above all in the training data. Many datasets used to “teach” AI contain demographic notes that should not be relevant for diagnosis. But algorithms, programmed to look for correlations, end up giving too much weight to factors such as age, gender or ethnicity, to the detriment of much more meaningful clinical indicators. It’s the paradox of technology: the more powerful it is, the more it risks picking up the wrong signals if it is not guided carefully.
A correctable problem (if you recognize it)
The good news, underline the researchers and relaunched by ANSA, is that these errors can be corrected. The proposed solution consists of a rebalancing of the parameters analyzed by the AIreducing the weight of demographic factors and strengthening that of elements directly linked to the disease. A discovery that has a double value: on the one hand it highlights the risks of an uncritical use of artificial intelligence in medicine; on the other it demonstrates that fairer models are possibleprovided that systems are designed with greater awareness.
Technology yes, but with responsibility
The message that emerges from the study is clear: theAI is not neutral by definition. It reflects the data with which it is trained and the choices of those who design it. «If we are more aware and attentive to the way we build artificial intelligence systems – explain the authors – we can develop fairer models». In an era where technology enters ever deeper into clinical decisions, the challenge is not just to innovate, but do so ethically and fairly. Because even an algorithm, if not controlled, can end up discriminating. And in medicine, this is not an acceptable mistake.