A ventilator allocated wrong is a ventilator someone else won't have
In the depths of the winter 2021 surge, when hospitals faced the impossible arithmetic of too many patients and too few resources, a Franco-American research consortium offered a rare gift: foresight. Owkin's AI-Severity Score, published in Nature Communications, learned to read lung CT scans with a precision that outpaced every existing clinical benchmark, giving overwhelmed medical teams the ability to see, before crisis struck, which patients would need the most. It is a reminder that in moments of collective suffering, the most humane act can sometimes be the algorithmic one — turning data into decisions that save lives.
- Hospitals in January 2021 were operating at the edge of collapse — beds full, ventilators scarce, and clinical staff forced to make life-or-death triage calls with incomplete information.
- The core tension was one of allocation: a ventilator given to the wrong patient at the wrong moment could cost another person their life, making accurate severity prediction not a convenience but a moral imperative.
- Owkin's AI-Severity Score attacked this problem by mining CT lung scans — already being taken on arrival — for patterns invisible to the human eye, trained across multimodal datasets and benchmarked against every published rival model, which it surpassed.
- The ScanCovIA consortium — drawing together Institut Gustave Roussy, Kremlin-Bicêtre APHP, CentraleSupélec, INRIA, and Owkin — demonstrated that pooling institutional data across borders could achieve what no single hospital working alone could.
- Looking beyond the immediate crisis, Owkin was simultaneously building epitope prediction models to identify the most stable, mutation-resistant parts of the coronavirus — tools designed to make future vaccines more durable against emerging variants.
- The trajectory points toward precision medicine at scale: machine learning not merely reacting to disease, but anticipating it — matching the specific biology of each patient and each pathogen to the care most likely to matter.
By late January 2021, the vaccine rollout had begun, but hospitals were still overwhelmed. Beds were full, staff exhausted, and ventilators in short supply. Into this crisis, a French-American startup called Owkin published a machine learning model in Nature Communications designed to do something hospitals urgently needed: predict which COVID-19 patients would deteriorate, and how quickly.
The tool, called the AI-Severity Score, worked from CT scans already being taken when patients arrived — images radiologists had been reading for months. The model was trained on multimodal data and, when tested against existing severity prediction benchmarks, outperformed them all. The research emerged from the ScanCovIA consortium, a collaboration between Institut Gustave Roussy, Kremlin-Bicêtre APHP, Owkin, and the Digital Vision Center of CentraleSupélec and INRIA — institutions combining data and expertise to solve a problem no single hospital could tackle alone.
The stakes were not abstract. Knowing which patients will worsen allows hospitals to direct ventilators and intensive care to those who will die without them, rather than those who will recover on oxygen alone. Better prediction means better triage, and better triage means lives saved.
Owkin was also thinking further ahead. The company was developing machine learning models to identify epitopes — the parts of the coronavirus most likely to provoke a strong immune response and, crucially, least likely to mutate away. If stable epitopes could be identified faster, vaccines could be designed to hold up longer against new variants, a concern already pressing as the first mutations began to emerge. The same techniques carried implications for cancer research, where understanding immune recognition of tumor cells is equally vital.
Underneath both efforts was a single ambition: a deeper, more precise understanding of how the immune system works — and how medicine might finally be tailored to the specific biology of each patient and each disease.
By late January 2021, vaccines were rolling out across the world, but hospitals were still drowning. Beds were full. Staff were exhausted. Ventilators were scarce. In this moment of crisis, a French-American startup called Owkin published a machine learning model in Nature Communications that could do something hospitals desperately needed: predict which COVID-19 patients would get sickest, and how fast.
The model, called the AI-Severity Score, works from something hospitals already had on hand—CT scans of the lungs taken when patients arrived. Radiologists had been looking at these images for months. Now a machine could look at them too, and extract patterns humans might miss. The model was trained on multimodal datasets, meaning it learned from multiple types of data at once, and when researchers tested it against existing severity prediction tools, it outperformed them all.
Why this matters is not abstract. When a hospital knows which patients will deteriorate, it can move resources where they're needed most. A ventilator allocated to someone who will recover on oxygen alone is a ventilator not available to someone who will die without it. Better predictions mean better triage, which means lives saved. The research came from a consortium called ScanCovIA, a collaboration between Institut Gustave Roussy, Kremlin-Bicêtre APHP, Owkin, and the Digital Vision Center of CentraleSupélec and INRIA—institutions pooling their data and expertise to solve a problem that no single hospital could solve alone.
But Owkin was thinking beyond the immediate crisis. The company was also building other machine learning models aimed at a different problem: finding the parts of the coronavirus most likely to trigger a strong immune response. These are called epitopes, and they matter because viruses mutate. The vaccines being distributed in January 2021 were effective, but nobody knew how long that would last. The flu requires new vaccines every year or two. COVID-19 might do the same.
If researchers could identify which epitopes were most stable—least likely to change as the virus evolved—they could design vaccines that would hold up longer against new variants. Machine learning could speed this up by narrowing the field of candidates to test in the lab, rather than forcing scientists to check every possibility by hand. This kind of epitope prediction has applications beyond vaccines too. The same techniques could be used in cancer research, where understanding how the immune system recognizes tumor cells is equally crucial.
What Owkin was really after was deeper: a better understanding of how the immune system works. The machine learning models were tools toward that end, part of a larger push in precision medicine to match treatments to the specific biology of each patient and each disease. In January 2021, with hospitals overwhelmed and variants beginning to emerge, that kind of precision felt urgent. The severity score could help today. The epitope models might help tomorrow.
Citas Notables
Hospital resource management, when done well, saves lives— Owkin research team (paraphrased from Nature Communications publication)
La Conversación del Hearth Otra perspectiva de la historia
Why does predicting severity from a CT scan matter more than, say, a blood test or a clinical score a doctor could calculate by hand?
Because a CT scan shows the actual damage the virus has done to the lungs—the inflammation, the scarring, the distribution of injury. A machine can see patterns across thousands of scans that no individual radiologist could hold in their head. And it's fast. In a hospital at capacity, speed is survival.
But hospitals were already using CT scans. What changed?
The machine learning model turned images into predictions. Before, a radiologist would describe what they saw. Now the model could say: this patient will need a ventilator in 48 hours. That's actionable. That's resource planning.
The article mentions epitope prediction for vaccines. Why is that connected to severity prediction? They seem like different problems.
They're both about understanding the virus better. Severity prediction tells you who's in danger now. Epitope prediction tells you how to defend against the virus later. Owkin was working on both because the pandemic wasn't over—it was just beginning to shift.
What happens if the virus mutates faster than new vaccines can be made?
That's the fear. If you can identify the parts of the virus that don't mutate, you can build vaccines that stay relevant longer. Machine learning can screen thousands of possibilities in weeks instead of months. It's not a guarantee, but it's better odds.
Does this model work for other diseases?
The epitope prediction does—cancer research is already using similar techniques. The severity score is specific to COVID-19 right now, but the approach could apply to any disease where imaging predicts outcomes. Pneumonia, for instance. Or sepsis.