Machine learning algorithms applied to data from the Australian IPF registry have identified the key risk factors for IPF progression.
Dr John Mackintosh told the Australasian Rare Lung Disease Conference 2021 that once validated, the variables of interest could hopefully be input into a user-friendly online IPF prediction tool to help inform clinical decision making.
“That might allow us to counsel patients a little bit better – to say, you are definitely likely to progress in the next 12 months, we really should be starting antifibrotics, particularly in those patients …who may have had a mild or fairly benign trajectory to date,” he said.
Dr Mackintosh, a respiratory and lung transplant physician at the Prince Charles Hospital, told the conference that the research aimed to build upon the GAP (Gender, Age, Physiology) model.
“At least in my practice I very rarely implement the GAP model, even in the lung transplant setting. I think that’s because it doesn’t necessarily help us to inform what is going to happen to the particular individual in front of us.”
“The other thing we know about the GAP model is it doesn’t help to inform us what is going to happen with a person’s FVC over a period of time. It doesn’t matter whether you are in GAP stage 1 or III, the rate of change in the FVC is the same and so it doesn’t necessarily tell the individual what is going to happen to them over 12 months.”
The study applied a suite of machine learning algorithms to interrogate about 250 variables in the data set from the IPF registry.
The study found that at a first presentation, % predicted DLCO, % predicted FVC, WHO functional class, the presence of clubbing, SGRQ total score, SGRQ breathless at home and SGRQ side effects from medication, had an accuracy of up to 67-70% in predicting a combined outcome of a significant drop in FVC or DLCO at 12 months, transplant or death.
“When a patient then presents for a follow up visit at any time point, we get a little more accurate in predicting that composite endpoint.”
The variables that become significant are % predicted DLCO, cough severity on a VAS, a couple of measures from the SGRQ including discomfort in public or side effects from medication, the presence of pulmonary hypertension, and then change in DLCO and change in FVC.”
Dr Mackintosh said a range of lung function variables, along with “slightly unusual” variables such as heartburn when bent over and SGRQ breathless when talking, were significant in predicting FVC/DLCO decline alone at a first presentation.
At a subsequent presentation, change in FVC and change in DLCO variables had an 85% accuracy in predicting future change in those parameters.
“I don’t think that is surprising to anyone,” he said. “What’s interesting is that in previous studies, changes in those variables haven’t necessarily gone on to predict subsequent change.”
He said the accuracy of the models were even better – getting close to 90% – in predicting death or transplant.
Lung function parameters, BMI and some individual measures from the SGRQ remained relevant at a first presentation and CT honeycombing appeared as a significant variable for ongoing presentations.
Dr Mackintosh’s presentation was awarded best clinical abstract at the conference.