top of page

Machine Learning Model Pinpoints Patients for Intubation Before They Crash

Many COVID-19 patients admitted to the hospital seem to do fine, until they crash.

One of the more puzzling cohorts of COVID-19 patients are those who appear to be stable, but suddenly crash and become in urgent need of intubation. Dr. Collin Stultz, a biomolecular engineer and practicing cardiologist at Massachusetts General Hospital (MGH), is developing a machine-learning model to identify these patients before their prognosis is dire. Dr. Stultz will be presenting his model at the talk, “Predicting Respiratory De-compensation in Patients with COVID-19,” at the AI Cures Conference on September 29th.

At MGH, Stultz sees patients enduring “a very precarious course.” After days of doing well, some patients deteriorate rapidly over a span of 12 to 24 hours. “They end up intubated in the intensive care unit,” Stultz says.

Simple models, which leverage data from routine lab tests and patient demographics, can only marginally discriminate between patients who are at high risk of needing to be intubated and those who will not. Although such models perform better than clinical intuition alone, “nevertheless we believe we can do better,” Stultz says.

To identify patients at risk of being intubated, Stultz’s lab designed a machine-learning model that incorporates data from both routine lab tests and electrocardiograms (ECGs). An ECG, which provides information about the heart’s electrical system, typically records between 250 and 500 points of data every second. A human cannot process that much information and come to any meaningful conclusion, “but a computer can,” Stultz says. The algorithm combines the ECG and lab data in impossibly complex ways, yielding a much more accurate picture of risk. The end result is a risk score, which a physician can interpret and use to determine the best therapy for a given patient.

Since the pandemic began, Massachusetts General Hospital has gathered a database of over 500 COVID-19 patients, nowhere near enough data to train a deep-learning model from scratch. To sidestep this problem, Stultz first trains a model on 7 million ECGs obtained from patients admitted with other diagnoses. This model is designed to estimate markers that indicate increased sympathetic tone – a harbinger of adverse outcomes. “You combine the information from 7 million patients, and you just have to fine-tune it to the task that is most relevant to the 500 patients with COVID-19,” Stultz says.

Stultz’s lab began developing this algorithm, before the pandemic, to predict adverse outcomes in patients admitted with congestive heart failure. However, whether it’s COVID-19 or another disease, the algorithm uses ECG data to predict increased cardio-pulmonary stress. The hope is that the model can detect very subtle changes in the sympathetic nervous system before those changes manifest as physical symptoms. The model can identify patients in distress before a physician becomes aware that there is a looming crisis, allowing an intervention that may ameliorate a poor patient outcome. The work is ongoing, but so far, promising.

Eventually, Stultz hopes to set up software that can allow doctors to easily upload ECG and lab data for any patient. The software would be made public, making the model accessible to clinicians at any hospital in the world.

bottom of page