New Machine-Learning Models Can Predict 6-Month Cancer Mortality
A new study shows that machine-learning models can predict six-month mortality for cancer patients, which has the potential to aid clinicians with prognosis.
Source: Getty Images
By - Researchers have validated and developed machine-learning (ML) models that can predict six-month mortality for patients with advanced solid tumors who are considering a new line of therapy (LoT).
A new study published in JCO Clinical Cancer Informatics evaluated whether it was possible to utilize EHR data to develop a risk model that can predict whether patients with an advanced solid tumor will be deceased within six months of beginning a new LoT.
The researchers state accurate estimation and communication of prognosis allows clinicians to refer patients for palliative and hospice care earlier while increasing the likelihood that patients will receive end-of-life (EOL) care — that is, the care received in the six months before death — consistent with their wishes.
According to the study, clinicians often overestimate survival, indicating that an accurate ML prognostic tool may improve decision-making by clinicians and their patients. To develop their models, the researchers gathered data from the University of Utah Health (UHealth) enterprise data warehouse.
EHR data from inpatient or outpatient UHealth and Huntsman Cancer Institute (HCI) care settings were pulled alongside HCI hospital-based cancer registry (HCI-CR) information to identify the study cohort and determine predictive features.
The study cohort included patients who had advanced solid tumors and received care within UHealth’s network. Of these, 4,192 patients with at least one LoT with a treatment decision point (TDP) were selected for the final study population. After evaluating cancer prognostication literature and evaluating the previously gathered data, the researchers found 111 relevant predictive features to include in their full model. They also created two additional limited models with and without cancer-related features, which lowered the number of features to 45 and 23, respectively.
Overall, predictive performance was high and similar among all models. Of the 7,056 TDPs experienced by the study population, 8.5 percent were flagged as having a low chance of survival. Among the patients whose TDPs were flagged, 66 percent died within six months.
Some of the most relevant predictive features across the models were albumin, pain score, time to next treatment (TTNT), time since index date, and age. The similar performance across models, regardless of cancer-related features, may indicate a common pathway at EOL for all advanced solid tumor patients, within which features unrelated to cancer become more important.
The researchers suggest that since performance was similar for all three models, the 45-feature model may be more easily scaled, implemented, and interpreted by health professionals for clinical use.
With the growing use of AI in healthcare, numerous studies have utilized predictive analytics to enhance cancer care.
A recent study published in Nature showed swarm learning models could be used to predict cancer biomarkers from standard histopathology images. The researchers found that not only could the models accurately predict biomarkers, but that they also outperformed conventional AI methods used for biomarker prediction. The study suggested that swarm learning models could be used in the future to address the most common issues associated with traditional predictive models.
