Machine studying method utilizing step rely predicts hospitalization throughout radiation remedy

October 26, 2022

2 minute learn

Disclosures: Friesner doesn’t report any related monetary info. Please see the research for related monetary info from all authors.

We’ve got not been capable of course of your request. Please attempt once more later. In case you proceed to have this drawback, please contact

A synthetic intelligence mannequin appeared to foretell the chance of unplanned hospitalizations throughout chemoradiotherapy amongst a cohort of sufferers with varied varieties of most cancers.

The outcomes, offered on the annual assembly of the American Society for Radiation Oncology, confirmed that the mannequin, which used day by day step counts measured utilizing wearable gadgets as a proxy for monitoring affected person well being, supplied physicians with a real-time methodology to ship customized care.

Researchers used daily step counts and other patient data to develop a machine learning model that can predict the likelihood of unplanned hospitalization during cancer radiation therapy.  Source: Adobe Stock
Researchers used day by day step counts and different affected person information to develop a machine studying mannequin that may predict the chance of unplanned hospitalization throughout most cancers radiation remedy. Supply: Adobe Inventory

Context and methodology

About 10% to twenty% of sufferers present process outpatient radiation remedy or chemoradiotherapy require acute care through an emergency room go to or hospitalization throughout their remedy. These unplanned hospitalizations can result in remedy delays and stress that may have an effect on scientific outcomes, in line with a information launch.

Portable devices allow steady and goal seize of patient-generated well being information exterior of the scientific setting, which minimizes journey and has the potential to have a extra lifelike and honest evaluation of an individual’s well being standing,” Isabella Friesner, scientific information researcher on the College of California, San Francisco, mentioned throughout the presentation. “Whereas wearable machine information has typically been used to encourage sufferers to extend bodily exercise, this research is exclusive in that we’re constructing machine studying fashions primarily based on step rely information to establish outcomes potential scientific outcomes, i.e. hospitalizations.”

Friesner and colleagues used information from three potential, single-institution trials of exercise monitoring in sufferers present process chemoradiotherapy. Every affected person acquired a health tracker and wore it earlier than and through chemoradiotherapy remedy.

“Earlier work from these trials confirmed a big affiliation between current step rely and danger of hospitalization, the place a discount in danger was related to a rise in day by day step rely,” Friesner mentioned.

The researchers used the information to develop and internally validate three machine studying approaches primarily based on day by day step counts throughout chemoradiotherapy remedy to foretell hospitalization charges, together with elastic net-regulated logistic regression, networks of rare-input neurons and the random forest.

The research included 214 sufferers (median age, 61 years) with 11 main most cancers websites, together with head and neck (30%) and lung (29%). The researchers grouped 70% of the sufferers right into a coaching set (n=151) and the remaining 30% right into a ready check set (n=63). For every affected person, the researchers analyzed information recorded by the health tracker from one week earlier than chemoradiotherapy till the primary hospitalization or the tip of remedy.

Major conclusions

The researchers reported an total first hospitalization fee of about 27%.

The outcomes of the three machine studying approaches confirmed that the elastic community method demonstrated the very best total efficiency, with an space beneath the curve (AUC) of 0.8, adopted by the neural community method with an AUC of 0.72 and random forest with an AUC of 0.65.

“We then took probably the most environment friendly mannequin, the elastic web, and wished to see the significance of the number of steps traits had been as much as the efficiency of the mannequin,” Friesner mentioned. “We created a mannequin with solely the step rely and a mannequin with solely the scientific options/with out the step rely.”

The outcomes confirmed that the mannequin that included the step rely considerably outperformed the mannequin with out the step rely (AUC, 0.8 versus 0.46).

Limitations of the research included the small information set of simply 214 sufferers, Friesner mentioned.

Future analysis

“These information underscore the significance of step counts for the predictive capacity of a machine studying mannequin,” Friesner mentioned. “Machine studying will be utilized to day by day exercise monitoring information to foretell hospitalization.”

The researchers plan to conduct extra rigorous validation of the algorithm on the NRGF-001 trial, which is able to randomly assign sufferers present process chemoradiation for lung most cancers to remedy with or with out day by day step rely monitoring. Further research are additionally deliberate to look at further parameters collected by wearable gadgets, similar to coronary heart fee, and their usefulness within the clinic.


  • AI mannequin utilizing day by day step rely predicts unplanned hospitalizations throughout most cancers remedy (press launch). Out there at: Printed October 24, 2022. Accessed October 24, 2022.
  • Friesner I, et al. Summary 132. Introduced on the American Society for Radiation Oncology Annual Assembly; October 23-26, 2022; New Orleans.

Leave a Comment