A machine learning model that differentiates between lupus and other rheumatic diseases might not improve diagnosis as designed, but could help refine SLE classification criteria.
A machine learning model that differentiates between lupus and other rheumatic diseases might not improve diagnosis as designed, but could help refine SLE classification criteria.
Diagnosing lupus can be difficult, considering skin manifestations of SLE appear similar to scleroderma and psoriasis, and symptoms can be subtle.
A tool designed to aid diagnosis and enable prompt treatment of this typically unpredictable disease could help reduce disease flares and prevent irreversible organ damage.
The new model evaluates a combination of common clinical and serological features of lupus derived from three sets of disease classification criteria and a large discovery dataset of rheumatology patients.
According to its developers, the model shows an “excellent combination of sensitivity and specificity” for detecting SLE against competing rheumatologic diseases.
But any machine learning algorithm is only as good as the data it was trained on, and further work will be required to independently validate the model’s performance in more diverse populations, since it has only been trained and tested using retrospective cohorts thus far.
The team designing this tool extracted data on roughly 800 people diagnosed with either lupus or another rheumatic disease at two Greek hospitals between 2005 and 2019, to construct, train and compare a pair of machine learning algorithms.
“We applied machine learning on panels of clinical features aiming to construct a model that can accurately detect SLE against competing rheumatologic conditions,” the researchers wrote in their paper.
The best performing model, called SLE Risk Probability Index (SLERPI), considered 14 clinical and serological features, including one not listed in current classification criteria, interstitial lung disease, as an additional predictor.
It classified patients, with varying certainty, into groups according to their likelihood of having lupus: unlikely, possible, likely and definite SLE.
However, the algorithm had greater accuracy when operating as a binary model to predict whether or not a person had lupus, and not another rheumatological disease.
It also showed a high sensitivity (over 92%) for identifying early and severe forms of lupus, and certain disease subtypes.
This was estimated using a second retrospective cohort of over 500 lupus patients and nearly 150 controls, consecutively registered in the clinics’ medical records.
While the machine learning model is good in principle, the algorithm is limited by design, said Dr Chris McMaster, a data scientist and rheumatology fellow at Austin Health in Melbourne.
The dataset used to train the algorithm featured patients diagnosed at rheumatology clinics by experienced clinicians using established classification criteria for SLE, Dr McMaster said, “so that determination [of lupus-related features] has already been made by a dermatologist or rheumatologist.”
This means the tool could not be readily applied in other settings, such as GP clinics and emergency departments where clinicians determine whether to refer patients to specialist care.
Dr McMaster also said the tool computed evidence collected over a lifetime of a person having lupus, rather than processing the information a clinician might have at hand to make a diagnosis.
Where the tool could be useful instead, Dr McMaster said, is in refining the current classification criteria for lupus, with the algorithm providing a data-driven approach to classify disease.
“To this end, establishing a firm diagnosis and treatment plan still remains at the judgement of experienced physicians,” the researchers concluded.
