AI model predicts 5-year Stevens-Johnson syndrome ocular symptom prognosis
Click Here to Manage Email Alerts
Key takeaways:
- There is no standardized prediction system for Stevens-Johnson syndrome, which is rare.
- The model used included 216 ocular segment images from patients with chronic SJS and 50 from healthy patients.
An artificial intelligence system predicted the severity of ocular symptoms in Stevens-Johnson syndrome with toxic epidural necrolysis 5 years after diagnosis, according to a study published in Allergy.
These prognoses were comparable with the inferences of an experienced physician, Mayumi Ueta, MD, PhD, associate professor, department of frontier medical science and technology for ophthalmology, Kyoto Prefectural University of Medicine, and colleagues wrote.
Stevens-Johnson syndrome with toxic epidermal necrolysis (SJS/TEN) can damage the cornea and cause vision loss, the researchers said. But because it is rare and lacks a standardized prediction system, they continued, treatment greatly depends on clinical experience.
The researchers then built and evaluated the accuracy and feasibility of a prognosis prediction model for SJS/TEN treatment for applications in clinical practice based on AI.
The model used anterior ocular segment images from 216 outpatients with SJS/TEN who had not had any ocular surface operations and from 50 patients with normal corneas and no SJS/TEN as well as the inferred severity of conjunctival invasion of the cornea and related clinical data.
The researchers began by creating a classifier model that infers the status quo severity of the SJS/TEN ocular surface findings with diagnostic discrete severity information of the conjunctival invasion of the cornea.
Areas under the receiver operating characteristic (ROC) included 0.92 for severity class 0, 0.91 for severity class 1, 0.86 for severity class 2 and 0.98 for severity class 3, indicating that the model can accurately classify severities based on anterior ocular segment images.
The researchers then adapted the severity inference model to infer severity at status quo as continuous values from 0 to 3 to create the prognosis prediction model with high granularity.
These continuous values enabled the researchers to create an AI model that predicted the severity of ocular conjunctival invasion more than 5 years after the first visit, they said, including inferred severity at the first and last visit in addition to the results of medical inspections and other data in their dataset. The latest inferred severity variables were the target variable.
The researchers said they confirmed the distribution of the latest severity in the training and holdout datasets were comparable and that both datasets covered the whole range of severity, indicating the model was not evaluated based on a selection of severity parts.
Next, the researchers compared an empirical model that mimicked an experienced physician’s prognosis with the performance of the AI models via coefficient of determination and mean absolute error metrics.
The performance of the best Light Gradient-Boosting Machine was comparable with the empirical model, the researchers said, indicating that patterns in the inference of the experts and in the dynamics of symptom progression were learned by the AI model.
Based on these findings, the researchers concluded that it is possible to create an AI system that can reproduce the inferences of an experienced physician and predict 5-year SJS/TEN ocular surface prognoses.
But although these prognoses may be useful in planning treatment for chronic ocular SJS/TEN, the researchers continued, additional studies of the feasibility of these models are warranted.
Perspective
Back to Top
Jodhbir S. Mehta, MBBS, PhD, FRCOphth, FRCS (Ed), FAMS, FARVO
These findings are interesting and, I would say, surprising, since the number of clinical photographs used is only approximately 216, and the ROC curves were very high, especially for the cases that had the worst outcomes. They are significant since they give a clinician the ability to prognosticate and alter their management.
SJS/TEN is a difficult disease to treat. It is rare but can have devastating mortality and morbidity. Early presentation and management are really important to avoid long-term sequelae, but despite that, outcomes can be poor. We have not used an AI algorithm to determine long-term outcomes, but I think this is a great and innovative idea and use of AI.
AI models like the one here will play a larger role in diagnosis and predicting diseases like this, especially if we can get larger collaborative databases. I am surprised at how good the ROC is with just clinical photos and clinical history. It would be good to validate this result on datasets from other countries as well.
The next step in this research should include validation of the results and software in other ethnic groups. In addition, increasing the size of the dataset using artificially generated images can help improve the ROC.
Collapse
Click Here to Manage Email Alerts