machine learning algorithm that uses data from electronic medical records (EMRs), can accurately predict early mortality risk in patients with cancer who are going to undergo chemotherapy indicates a retrospective cohort study.
"New chemotherapy is a critical event in the trajectory of the disease of cancer, and objective predictions of short-term mortality in this period may be helpful to physicians and patients in several ways," - said senior author Ziad Obermeyer, MD, MPhil, Brigham and Women's Hospital and Harvard Medical School, Boston, Massachusetts, and colleagues.
"This model has worked well in different types of cancer, race, gender, and other demographic data" - they write.
"The estimates were accurate for chemotherapy regimens with palliative and curative purpose, for patients with cancer early and distant stage, as well as for patients receiving circuit clinical trials, introduced a few years after learning model," - they added.
Further research is needed to determine whether it is advisable to apply this algorithm in other clinical conditions, they warn.
The study was published July 27 in JAMA Network Open.
For their study the team analyzed data from the EMR for all patients undergoing chemotherapy at Dana-Farber / Brigham and Women's Cancer Center in Boston, Massachusetts.
"We identified 26,946 patients who initiated a 51,774 discrete chemotherapy from 2004 to 2014", - wrote the researchers.
The average age of the group was 58.7 years, 61.1% - women, 86.9% - white.
During the start of chemotherapy, 59.4% were on a distant stage disease.
Reporting on the results of the validation model only, and not by derivation model on which it was founded, the researchers say that the overall 30-day mortality rate was 2.1% among the 9114 patients included in the validation set.
"The model accurately predicted 30-day mortality for all patients regardless of the purpose of chemotherapy," the authors argue.
Among patients undergoing palliative chemotherapy - for whom the prognostic evaluation will be particularly important, the model also works well.
The researchers also used the model to rank individual patients with palliative chemotherapy for early risk of mortality at 30 days.
In this subgroup of patients, they found that 30-day mortality was 22.6% among the highest deciles of predicted risk of risk versus 0% for patients in the decile with the lowest level of risk.
The team then used the model to predict the risk of death by 180 days.
Among all patients included in the validation set, the overall mortality for 180 days was 18.4%; among patients who underwent palliative chemotherapy, mortality within 180 days was higher - 27.9%.
"The model predictions of 30-day mortality were accurate predictors of mortality at 180 days," - the researchers note.
Again, for ranking in the highest decile of the risk of mortality at 180 days was 74.8%, compared with 0.2% among patients in the decile with the lowest level of risk.
The researchers also used the model to patients with distant stage disease. In this subgroup, the average 30-day mortality rate was 2.9%.
However, again, the risk of mortality in 30 days was significantly higher - at 22.7% among patients with decyl with the highest level of risk compared to 0% of the decile with the lowest level of risk.
Even when the model was used for the experimental regimens initiated from 2012 to 2014, the accuracy of the predicted model validation was very high, while the AUC 0.942, the researchers note. This is despite the fact that the model used to train the algorithm, not subjected to these new schemes.
Analyzing only patients with remote stage, the authors compared the effectiveness of their model with two external assessments of mortality - these randomized controlled trials (RCTs) and surveillance data registers, Epidemiology and End Results (SEER).
The researchers note that the data from the SEER registry trials and are often used by doctors for mortality projections.
"Overall AUC for RCTs was 0.555 ... compared with 0.771 for the model-based estimates", - reported the team.
Model predictions similarly exceeded SEER ratings for a 1-year mortality rate for the same patients.
The study authors note that, to be useful, predictive models should help the physicians to make fundamental decisions in everyday clinical practice.
They assume that machine learning algorithm, such as his own, that can identify cancer patients with a high risk of early death, can help in decision-making and patient doctor initiation of chemotherapy and pre-planning assistance. "