AI model predicts the likelihood of unplanned hospitalization during radiation treatments for cancer

AI model predicts the likelihood of unplanned hospitalization during radiation treatments for cancer

A synthetic intelligence (AI) mannequin developed by researchers can predict the chance {that a} affected person might have an unplanned hospitalization throughout their radiation therapies for most cancers. The machine-learning mannequin makes use of every day step counts as a proxy to watch sufferers’ well being as they undergo most cancers remedy, providing clinicians a real-time technique to supply personalised care. Findings will probably be introduced immediately on the American Society for Radiation Oncology (ASTRO) Annual Assembly.

An estimated 10-20% of sufferers who obtain outpatient radiation or chemoradiation remedy will want acute care within the type of an emergency division (ED) go to or hospital admission throughout their most cancers remedy. These unplanned hospitalizations generally is a main problem for individuals present process most cancers remedy, inflicting remedy interruptions and stress which will affect medical outcomes. Early identification and intervention for sufferers at greater threat of issues can stop these occasions.

When you can anticipate a affected person’s threat of unplanned hospitalization, you’ll be able to change the way you help them via their most cancers therapies and cut back the chance that they are going to find yourself within the ED or hospital.”

Julian Hong, MD, Examine’s Senior Creator

Julian Hong is an assistant professor of radiation oncology and computational well being sciences on the College of California, San Francisco (UCSF), the place he additionally serves as Medical Director of Radiation Oncology Informatics.

Dr. Hong’s crew beforehand demonstrated {that a} machine studying algorithm utilizing well being knowledge resembling most cancers historical past and remedy plan may establish sufferers at greater threat of ED visits throughout most cancers remedy, and that further surveillance from their suppliers diminished acute care charges for these sufferers.

For the present examine, he and Isabel Friesner, lead creator and a medical knowledge scientist at UCSF, collaborated with Nitin Ohri, MD, and colleagues at Montefiore Medical Heart in New York to use machine studying approaches to knowledge from wearable shopper gadgets. Dr. Ohri and his crew beforehand collected knowledge from 214 sufferers in three potential medical trials (NCT02649569, NCT03102229, NCT03115398). In every of those trials, members wore health trackers that monitored their exercise over a number of weeks whereas they obtained chemoradiation remedy. Trial members had several types of major cancers, mostly head and neck (30%) or lung (29%) most cancers.

Step counts and different knowledge from these sufferers’ information have been used to develop and take a look at an elastic net-regularized logistic regression mannequin, a sort of machine-learning mannequin that may analyze a considerable amount of complicated info. The purpose of their mannequin was to foretell the chance {that a} affected person could be hospitalized within the subsequent week, based mostly on their earlier two weeks of information.

Researchers first created the mannequin by analyzing how properly totally different variables predicted hospitalization, utilizing knowledge from 70% of the trial members (151 individuals). Potential predictors within the mannequin included affected person traits (e.g., age, ECOG efficiency standing), in addition to exercise knowledge measured earlier than and through remedy. Along with every day step totals, the researchers computed different metrics, resembling relative modifications to an individual’s week-by-week averages or the distinction within the minimal and most variety of steps every week.

The analysis crew then validated the mannequin utilizing the remaining 30% of sufferers (63 individuals). The mannequin that built-in step counts was strongly predictive of hospitalization the next week (AUC = 0.80, 95% confidence interval [CI] 0.60-0.90), and it considerably outperformed the mannequin with out step counts (AUC = 0.46, 95% CI 0.24-0.66, p<0.001).

“The step counts instantly previous the prediction window ended up being typically extra predictive than medical variables. The dynamic nature of the step counts, the truth that they’re altering day-after-day, appears to make them a very good indicator of a affected person’s well being standing,” stated Dr. Hong.

The highest predictive variables within the mannequin included step counts from every of the previous two days, in addition to the relative modifications in most step depend and step depend vary over the previous two weeks.

Using dynamic knowledge differentiates this mannequin from these based mostly on medical knowledge like efficiency standing and tumor histology. “One of many distinctive components of this mannequin is that it is designed to be a working prediction,” defined Ms. Friesner. “You may run the algorithm on any given day and have an thought of a affected person’s threat stage one week out, providing you with time to supply that further help they want.”

This extra help is vital to decreasing hospitalizations, defined Dr. Hong, whether or not it is scheduling extra frequent follow-ups, altering one thing concerning the affected person’s remedy plan or one other personalised method. “The core of what works is that that is an additional touchpoint for a health care provider to see a affected person. It provides the affected person reassurance to know that we’re watching out for them.”

“As extra individuals start to make use of wearables, the query of whether or not the information they’re gathering could possibly be helpful arises. Our examine exhibits there may be worth in having our sufferers acquire their very own well being knowledge throughout their on a regular basis lives, and that we will use this knowledge to then monitor and predict their well being standing,” added Ms. Friesner.

Subsequent steps for the investigators embody a extra rigorous validation of the algorithm on the NRGF-001 trial (NCT04878952) led by Dr. Ohri, which can randomize sufferers present process CRT for lung most cancers to remedy with or with out every day step depend monitoring. Physicians of sufferers on the step depend arm will obtain output from the mannequin all through the remedy course of.

The researchers are additionally planning different research to look at further metrics collected by wearable gadgets, resembling coronary heart charge, and their utility within the clinic.

“Wearable gadgets and patient-generated well being knowledge are nonetheless comparatively new phenomena, and we’re nonetheless studying how they are often helpful. What different info can we receive from the numerous sensors in our lives? How can these metrics complement one another and work with different forms of knowledge, like digital well being report knowledge? Completely different datapoints may work for higher for various sufferers,” stated Ms. Friesner.

Following the widespread adoption of telemedicine and distant care over the previous a number of years, the necessity for distant monitoring by way of affected person gadgets may improve. Clinics and policymakers ought to maintain entry to those gadgets in thoughts as they develop in recognition, stated Dr. Hong.

“One of many challenges when working with real-world wearable knowledge are the financial and racial disparities that affect who owns gadgets that may seize the sort of knowledge. I feel it is essential to develop instruments which can be helpful for the clinic but additionally which can be accessible to a wider vary of sufferers.”


American Society for Radiation Oncology

Leave a Reply

Your email address will not be published. Required fields are makes.