AI diagnostics tools & oncology: predicting patient outcomes

AI French-American biotech start-up Owkin recently published new research in the European Journal of Cancer that shows how AI can predict the survival of cancer patients by looking at clinical data and radiology images.

Conducted over a period of three years, in conjunction with leading European cancer centre Gustave Roussy, the PULS-AI project aims to improve cancer care and help doctors better identify those patients most likely to benefit from treatment. And it now looks set to move from mere academic research project to tangible diagnostic product.

With a team of approximately 30 people – working with images from radiology, digital pathology, and oncology – the next step is to become a regulated medical device. Thanks to the radiologists at Gustave Roussy, the Owkin PULS-AI project has resulted in the creation of “a unique database of manually annotated radiology images” that includes 1,147 ultrasound images annotated with lesions delineation, in addition to 632 thorax-abdomen-pelvis CT scans fully annotated with some 9,516 lesions.

Machine learning in biology

Co-founded six years ago by Dr Thomas Clozel and Gilles Wainrib, a pioneering specialist in machine learning in biology, Owkin became a ‘unicorn’ start-up (valued in excess of $1 billion) in November 2021, after a $180 million investment from Sanofi.

Antoine de Sagey, director of diagnostics tools at Owkin, spoke with pharmaphorum about PULS-AI’s patient outcome prediction, citing the company’s earlier deep learning tool MesoNet’s success in predicting treatment outcomes in patients with the aggressive cancer, malignant mesothelioma (MM).

Although at the time of interview Dr de Sagey didn’t disclose the full seven cancers that PULS-AI has been able to predict patient outcome for through identification of biomarkers, he did explain the main facets of the technology and its findings.

Generally, Owkin machine learning technologies connect the dots between raw medical data and ultimate patient outcomes, but PULS-AI is the first AI diagnostics tool to be approved for use in assessing radiology and diagnostics, de Sagey said.

Identifying high- and low-risk groups

Using the Kaplan-Meier method – a non-parametric statistic for estimating survival function from lifetime data, or the fraction of patients living post treatment in a certain timeframe – the PULS-AI model is able to accurately stratify patients into high-risk and low-risk groups, predicting the outcome of treatment by utilising clinical data and radiology images.  Owkin trained a model with deep learning on CT scans and ultrasound images, as well as clinical data from over 600 patients across 17 different treatment centres in France.

Its PULS-AI results have now shown that “AI is able to extract relevant information from radiology images and aggregate data from different modalities to build powerful prognostic tools that could improve therapeutic decisions.” It holds promise for “integrating AI algorithms in the radiologist’s workflow to automate and homogenise this time-consuming annotation process”.

Focused on three main categories, the AI diagnostics tool helps pathologists in the identification of tumours and concerns surrounding those tumours. It also enables providers to connect patients with the correct treatments for them as individuals, based on biomarkers. Used in the pre-screening process that looks for specific proteins and genomic detail, the AI allows for accelerated diagnosis and a more targeted therapy.

Genomics and patterning in oncology

Genomics are a “big piece” of oncology, such an analysis looking for DNA changes or mutations, and AI essentially assists in sorting through and interpreting gathered genomic data. Thus, an AI diagnostic tool can be used to identify new drug candidates, both accelerate and de-risk clinical trials, and improve patient outcomes through ameliorated diagnostic solutions.

Relying on machine or deep learning, an AI’s algorithms identify patterns in data, compared with inputted data. In deep learning, the AI is formed from a series of artificial neural networks which operate like a human brain, assessing layers of data in order to translate all of it, via mathematical function, into new information (a ‘feature’), and so on.

Deep learning AI requires less human input and, as exhibited in the latest Owkin-Gustave Roussey trial, can reduce human working hours and the time it takes to analyse the masses of data and identify similarities and differences.

Phenotypic outcomes and patient prognoses can thus be more expediently predicted, and diagnostic and drug targets ascertained. What an AI diagnostic tool like PULS-AI does is “find the future patterns” and highlight the relapse risks and overall outcome prediction for the patient. In Dr de Sagey’s opinion, that’s pretty “unbelievable and crazy”.

From classic techniques to tomorrow’s AI outcomes

Classic techniques and practices traditionally, and by necessity, look at multiple slides from biopsies under microscopes. With AI, this process is fast-tracked to the extent that a patient outcome can be predicted from just one slide. Dr de Sagey did caveat that even more “specificity and sensitivity” are yet needed to be on a par with classic techniques, however.

Nonetheless, as Kathryn Schutte, senior data scientist and biomarker and target ID solutions manager at Owkin said, the mission is for PULS-AI to serve as “a tool that could provide therapeutic decision-making assistance to clinicians” at some point promisingly soon.

As AI is in its past, so it is very much in Owkin’s – and oncology’s – future, it seems.

The post AI diagnostics tools & oncology: predicting patient outcomes appeared first on .