Artificial Intelligence programs can outperform performance status-based scoring tools for predicting short-term mortality in advanced cancer patients, an Australian-led study suggests.
By applying machine learning to analyse clinical text in oncology letters from initial consultations, researchers were able to develop a model that was more precise in estimating patient 12-month survival rates compared to the prognostic estimates provided by oncologists based on ECOG performance status.
The researchers say the success of their computer program highlights the potential for automated AI applications that mine text in electronic medical records to glean other insights into patient groups.
Some 16 years of medical records from a major hospital system in New Zealand were fed into the model, which used machine learning tools to analyse each document line-by-line.
All up, so-called text mining was performed on the records of 4791 patients with stage IV cancer who had received treatment at the hospital from 2001 to 2017.
This enabled the program to “learn” various demographic and clinical details about each patient – up to and including death – and use that information to predict the survival of other patients.
Led by Dr Frank Lin of the Garvan Institute of Medical Research, Sydney, the team then put their model to the test over the following two years, comparing the accuracy of its predications versus those made by human oncologists using the Eastern Cooperative Oncology Group (ECOG) performance status (PS).
They found in 441 patients where clinical narratives from initial consultations were available, the AI program significantly outperformed the predictivity of PS.
And models were even more accurate when both predictions were combined, the researchers wrote in JCO Clinical Care Informatics (link here).
“Traditional prognostication tools operate by reducing complex symptomatology into summary scores (eg, PS),” they wrote.
“Such a reductionist approach, although simple to apply without a decision aid, invariably undermines the complexity of clinical scenarios.”
“We have demonstrated that, by exploiting the richness of clinical text with machine learning, more precise survival prediction can be obtained than with conventional PS-based scoring … supporting the leveraging of ubiquitous clinical text for augmenting PS-based prognostication.”
Based on the analysis, they noted there were a number of “recurring descriptors” correlating with survival, including referral patterns, mobility, physical functions and concomitant medications.
The order of medications ranked by risk of death broadly corresponded to the analgesic ladder. Methadone was associated with the worst prognosis, followed by fentanyl, levomepromazine, then morphine. On the other hand, amlodipine was associated with a good prognosis, as was budesonide and atorvastatin.
The authors added: “When applying clinical text to survival prediction, sophisticated black box models such as neural networks may offer better discrimination of outcomes.”
“However, our approach has the advantage of offering transparency in the feature selection process, producing interpretable results to allow covariates to be examined by experts for plausibility.”
“Analysis of EMR text has the potential to reveal patterns of observable clinical prognostic factors, and our findings of specific factors, such as mobility patterns and medical profiles, warrant further research to examine their values in prognostic modelling and supportive care,” they concluded