AI clinician trained to save humans from sepsis – and, er, let's just say you should stick to your human doctor

One out of three correct dosages ain't bad, right? Right?

Experts hope an artificially intelligent software system will help doctors tackle the deadly menace of sepsis in humans.

The technology uses reinforcement learning, an area of machine learning more commonly used for teaching bots to play games such as Go, Dota 2 and poker. In this case, however, instead of games, a software agent dubbed the AI Clinician toys with, er, human lives, in theory at least. All the experiments were conducted using just patient data, so no one was actually harmed in the process.

“The use of computer decision support systems to better guide treatments and improve outcomes is a much needed approach,” the team – based at Imperial College London, UK, and Harvard-MIT in the US – claimed in a paper published in Nature Medicine on Monday.

The team trained the software to assess patients from their records, and recommend an appropriate amount of intravenous (IV) fluids, a concoction of electrolytes and water administered directly to a patient’s veins, and vasopressors, a medication that increases blood pressure, to treat each patient suffering from sepsis. The AI Clinician can only recommend medication for folks during the early stages of sepsis, and the goal is to prevent the condition from worsening.

First, the team collected a large dataset containing 17,083 cases of patients with sepsis, and identified 48 variables per person that were relevant to treating the life-threatening complication. Some of these were demographic-based, such as age and gender, and other information included the amount of IV fluids and vasopressors received over four hour increments. The patient’s outcome – survival or death – was also included.

The goal was to identify the right combination of IV and vasopressors that maximized the chances of survival over 90 days for each patient. The training data was fed into the neural network, so that it could spot patterns and make recommendations for new patients based on their records. These suggestions were then compared to ones made by real doctors.

“On average, the AI Clinician recommended lower doses of intravenous fluids and higher doses of vasopressors than the clinicians’ actual treatments,” the paper said. For about 58 per cent of the time, the system predicted doses of vasopressor that were “very close” to the amount made by a real clinician. The results weren’t as strong for the IV fluid, agreeing with doctor's recommendations around 36 per cent of the time.

Unsurprisingly, the chances of a patient’s survival was highest where the model was most accurate in its recommendations compared to a real expert. The AI Clinician hasn’t been clinically tested yet, though the researchers hope that it can be used in real time in the future to guide doctors when recommending dosages. Individual patient data taken from electronic health records can be fed automatically into their algorithm to spit out a suggested dosage. Once it gets good at picking the right amounts, of course.

“Physicians will always need to make subjective clinical judgments about treatment strategies,” they warned. “Computational models can provide additional insight about optimal decisions, avoiding targeting short-term resuscitation goals and instead following trajectories toward longer-term survival.”

They also reckoned that even if the AI Clinician could only reduce mortality from sepsis by a small percentage, it would still save “several tens of thousands of lives” every year worldwide. ®