A machine learning algorithm that uses data from electronic health records (EHRs) can accurately predict early mortality risk in patients with cancer who are about to undergo chemotherapy, a retrospective cohort study shows.
"New chemotherapy is a critical event in the disease trajectory of cancer, and objective predictions of short-term mortality at this time could be useful to physicians and patients in several ways," senior author Ziad Obermeyer, MD, MPhil, Brigham and Women's Hospital and Harvard Medical School, Boston, Massachusetts, and colleagues observe.
"The model performed well across a range of cancer types, race, sex, and other demographics," they write.
"Estimates were accurate for chemotherapy regimens with palliative and curative intent, for patients with early- and distant-stage cancer, and for patients treated with clinical trial regimens introduced in years after the model was trained," they add.
Further research is needed to determine how feasible it is to apply this algorithm in other clinical settings, they caution.
The study was published online July 27 in JAMA Network Open.
For their study, the team analyzed data from EHRs for all patients undergoing chemotherapy at the Dana-Farber/Brigham and Women's Cancer Center in Boston, Massachusetts.
"We identified 26,946 patients who initiated 51,774 discrete chemotherapy regimens from 2004 through 2014," the investigators write.
The mean age of the group was 58.7 years, 61.1% were female, and 86.9% were white.
At the time of chemotherapy initiation, 59.4% had distant-stage disease.
Reporting on results only from the validation model and not the derivation model upon which it was based, the researchers state that the overall 30-day mortality rate was 2.1% among 9114 patients included in the validation set.
"The model accurately predicted 30-day mortality for all patients irrespective of chemotherapy intent," the authors affirm.
Among patients undergoing palliative chemotherapy — for whom prognostic estimates would be especially important — the model also performed well, with an area under the curve (AUC) of 0.924, they add.
The investigators also used the model to rank individual palliative chemotherapy patients for early mortality risk at 30 days.
In this subgroup of patients, they found that the 30-day mortality rate was 22.6% among those in the highest-risk decile of predicted risk vs 0% for patients in the lowest-risk decile.
Mortality at 180 Days
The team then used the model to predict mortality risk at 180 days.
Among all patients included in the validation set, the overall 180-day mortality rate was 18.4%; among patients who underwent palliative chemotherapy, mortality at 180 days was higher, at 27.9%.
"Model predictions on 30-day mortality were also accurate predictors of 180-day mortality," the investigators observe.
Again, for those ranking in the highest-risk decile, mortality at 180 days was 74.8%, compared to 0.2% among patients ranking in the lowest-risk decile.
The researchers also applied the model to those patients with distant-stage disease. In this subgroup, the mean 30-day mortality rate was 2.9%.
Again, however, mortality risk at 30 days was considerably higher, at 22.7%, among patients in the highest-risk decile compared to 0% among those in the lowest-risk decile.
Even when the model was used for experimental chemotherapy regimens initiated from 2012 to 2014, the predicted accuracy of the validation model was very high, at an AUC of 0.942, the researchers note. This is despite the fact that the model used to train the algorithm was not exposed to these new regimens, they add.
Comparison With Two External Sources
Analyzing patients with distant-stage disease only, the authors compared the performance of their model with two external sources of mortality estimates — randomized controlled trial (RCT) data, and data from the Surveillance, Epidemiology, and End Results (SEER) registry.
The researchers note that data from both RCTs and the SEER registry are often used by physicians to arrive at mortality predictions.
"The overall AUC for RCT estimates was 0.555...compared with 0.771 for model-based estimates," the team reports.
Model predictions similarly outperformed SEER estimates for 1-year mortality for the same patient populations.
Study authors point out that to be useful, predictive models must help physicians make pivotal decisions in daily clinical practice.
They suggest that a machine learning algorithm such as their own that can identify cancer patients at high risk for early mortality "may help to guide patient and physician decisions about chemotherapy initiation and advance care planning."
The study was supported by grants from the Office of the Director, the National Institute on Aging, and the Dana-Farber Cancer Institute. Coauthor Ravi B. Parikh, MD, has received personal fees from GNS Healthcare outside the submitted work. The other authors have disclosed no relevant financial relationships.
JAMA Network Open. Published online July 27, 2018. Full text
Δεν υπάρχουν σχόλια:
Δημοσίευση σχολίου