When researchers trained an artificial intelligence (AI) system to use radiological images to distinguish patients with COVID-19 pneumonia from those with other respiratory diseases, the machine found a logical – but faulty – short cut.
A radiologist would weigh up features of the images. “But the AI system learned to read the dates of the scan,” says Antonio Esposito, professor of radiology at Vita Salute San Raffaele University in Milan. The computer, he explains, simply put all patients who entered the hospital in 2020 into the COVID-19 category.
A new partnership between San Raffaele University and Microsoft aims to tackle such shortcomings and develop AI in health care to the point where it can reliably be introduced to improve patient care.
“This project removes walls between researchers’ and clinicians’ access to data from different hospital departments, and leverages the complementary IT and clinical expertise held by the two partners,” says Carlo Tacchetti, professor of anatomy at San Raffaele University.
Medical records, pathology reports and images are usually stored on separate databases in different hospital departments and labs, which can make it difficult to share the data they contain. The San Raffaele project will bring all relevant clinical data together in one system, combining new patient data with retrospective data from the past decade.
Together with Microsoft experts, doctors and researchers from San Raffaele University Hospital will use this information to tackle specific clinical AI challenges by developing tailored algorithms.
The AI technology will sift through all the data to identify patterns and make predictions about patient outcomes. In principle, the project will be able to share decision-making algorithms with clinicians elsewhere, who can then apply the machine learning technique to their own patient data to help guide their decision-making.
Opening the black box
In an important step, the decisions made by the AI will be transparent. “Unlike other projects, clinicians will be able to see the AI prediction pathway, allowing them to refine the algorithms, to remove any biases in the system and allow for machine and clinician learning from each other,” says Esposito.
This approach aims to avoid the kinds of biases that have plagued attempts to apply AI to health care, including “dangerous shortcuts sometimes hidden inside black-box AI solutions”, says Esposito.
Alongside the data collection and algorithm development, the university will open a dedicated school with Microsoft to offer PhD programmes that train computer scientists, engineers, clinicians and mathematicians to use AI in health care.
“With Microsoft’s help, we already have great minds working on this project from all over the world,” says Tacchetti. “We want to train the next generation of researchers in this field, alongside developing our AI systems.”
The hope is that the AI can develop “a better understanding of how a patient’s ethnicity, age, gender, occupation and geographical region can be linked with their condition”, Tacchetti says. “With our algorithms, expert knowledge could be instantly transmitted from specialist hospitals such as San Raffaele to clinicians in remote locations around the world that may have little experience of certain diseases.”