Article summary
An artificial intelligence program called DystoniaBoTXNet has identified areas of the brain that serve as biomarkers of treatment efficacy and identified patients who would benefit from botulinum toxin injections.
Two years after demonstrating that an artificial intelligence (AI) program could diagnose focal dystonias with extraordinary accuracy based solely on MRIs of the brain, Harvard scientists and her team discovered that another AI program could detect botulinum toxin. showed that it could predict with similar accuracy who would benefit from (BoTX) treatment.
Previous reports published in 2020 Proceedings of the National Academy of Sciences, used ‘deep learning’ or AI to analyze fine-structured neural networks in brain images. The system, called DystoniaNet, identified her three types of clinically diagnosed dystonia from healthy controls with 98.8% accuracy.
New report published November 28 Neurological Annalsused an AI program called DystoniaBoTXNet to identify eight cluster regions of the brain that act as biomarkers of treatment efficacy. identified with an overall accuracy of 96.3%, including 86.1% accuracy.
Senior authors of both papers said Neurology Today These two programs should help clinicians diagnose and treat focal dystonia faster, cheaper, and more accurately than other known methods.
“Currently, there is uncertainty as to whether patients will benefit. [BoTX] Therapeutic or not; Christina Simoyan, MD, MD, MD, Professor of Head and Neck Surgery, Otorhinolaryngology, Harvard Medical School, and General Eye and Ear Surgery, and Neuroscientist, Department of Neurology, Mass General Hospital said. “As a result of the uncertainty, many patients with potential benefit choose not to receive injections, while others go through multiple rounds without clinically significant benefit.” .”
“Patients and clinicians have no objective means available to predict the outcome of treatment prior to injection,” Dr. Simonyan said. Subjectivity may be introduced into the injector’s assessment and the patient’s perceived benefit, as “clinicians rely on post-injection symptomatic outcomes.”
“Our goal was to develop an automated, objective predictor that relies on the neuronal pathophysiology of dystonia, not just the outcome of symptoms,” said Dr. Simonyan.
Clinicians across the Mass General Brigham System are currently evaluating DystoniaBoTXNet as part of a larger clinical trial aimed at further clinically validating the algorithm, Dr. Simonyan said.
Survey details
AI programs of the type developed by Dr. Simonyan and her colleagues are designed to let data speak by automatically looking for important patterns without a clear direction. DystoniaBoTXNet applied what the authors called a ‘3D convolutional neural network architecture’ to raw structural brain MRI to identify neural network biomarkers of BoTX efficacy.
MRIs were from 284 patients with four types of focal dystonia: laryngeal dystonia, blepharospasm, cervical dystonia, and writer’s cramp. All patients were defined as either benefiting or not from BoTX treatment by asking each patient using a structured questionnaire based on a review of each patient’s medical information. it was done.
The program was initially applied to a so-called training set of 106 patients with laryngeal dystonia who benefited from BoTX treatment and 59 who received treatment but did not benefit. All patients had received injections at least 3 months prior to their MRI scan and were fully symptomatic. Although the patients were of similar age and had similar ages of onset and duration of dystonia, the patient who responded (7.1 years) was treated with her BoTX longer than the patient who did not (2.8 years). rice field. Each group used different scanner vendors, including Philips, Siemens, and GE.
DystoniaBoTXNet automatically identified eight clusters of interest: superior parietal lobule, inferior and middle frontal gyrus, middle orbital gyrus, inferior temporal gyrus, corpus callosum, inferior anteroposterior bundle, and anterior thalamic radiation. DystoniaNet has previously identified five of these eight regions. Three other clusters were recently reported to contribute to short- and long-term central responses of BoTX in patients with dystonia.
After DystoniaBoTXNet analyzed the MRIs of the first group, the results were tested in a first independent set of 29 laryngeal dystonia patients who benefited from BoTX and 15 patients who did not. A second independent set had other forms of focal dystonia, including 14 of hers with blepharospasm, 18 of hers with cervical dystonia, and 14 of hers with writer’s cramp. Overall, 38 patients benefited from her BoTX and 8 did not.
Finally, to assess outcomes in the untreated population, 29 patients with laryngeal dystonia underwent MRI scans and were analyzed by DystoniaBoTXNet.
In a training set of 165 patients, DystoniaBoTXNet showed an AUC (area under the curve) of 100% in discriminating between those who benefited from treatment and those who did not. This demonstrates the robustness of the algorithmic model.
Results for the first and second test sets of patients were equally accurate. This program took just 19.2 seconds to produce results.
Although the program predicted benefit in 3 patients, no such benefit was observed clinically. However, her two of these patients only received her one injection, suggesting that continuing additional injections would likely benefit.
In the final set of 29 treatment-naïve patients, the program predicted a median benefit probability of 94.6% in 23 patients and a benefit probability of 16.9% in the remaining 6 patients. Of her 7 patients who subsequently received BoTX treatment, 5 responded and 2 did not. DystoniaBoTXNet found that he was 100% accurate in his prediction of 5 of hers who experienced the effect, but also predicted the effect of 2 of hers who did not actually experience the effect. However, both of these patients had only received one of his injections, so Dr. Simonyan and colleagues concluded that these two were likely quasi-nonresponders, and that if they received additional injections concluded that it is likely to benefit from
Dr. Simonyan emphasized that no special training or technical knowledge of how the program works is required to integrate the program into clinical care.
“It is important to develop robust and validated AI tools. It’s equally important to ensure that it’s both translatable and practical,” she said.
“With DystoniaBoTXNet, the clinician clicks a few buttons to load the MRI into the program, enters the patient’s gender and age, and waits a few seconds for the odds of success. can be used together with a patient’s clinical assessment and medical history to develop a more advanced and individualized treatment paradigm,” she said.
Dr. Simonyan and colleagues have a patent on DystoniaBoTXNet that lists Dr. Simonyan as the inventor. The full details of her program’s algorithm have not been published, but interested scientists are welcome to contact her to request it for her research purposes.
Expert commentary
Neurologists familiar with the new paper interpreted the findings in different ways. Kelly A. Mills, M.D., Ph.D., associate professor of neurology at Johns Hopkins School, said: Director of the Department of Medicine and Movement Disorders, which studies and treats patients with dystonia. “Such a tool may help advise patients on whether to continue their efforts.”
Dr. Mills said the program could be particularly helpful for patients who have only had one or two injections and are unsure if they will benefit further. But for untreated patients, he said: [BoTX] If it is a first-line treatment for focal dystonia, we do not order injections based solely on the program’s prediction that the patient is unlikely to benefit. ”
But another neurologist said that while the AI program’s results were interesting, they were unlikely to change clinical practice in the near future.
said Brian D. Berman, MD, MS, FAAN, Professor of Neurology and Director of the Virginia Commonwealth Parkinson’s and Movement Disorders Clinic. University.
“However, the patient group in this study was dichotomous. [those who were] receive the benefit or not. The actual treatment for this disorder is not black and white. Patients can respond differently to toxin therapy, and poor responses are often due to the dose or injection strategy used. ”
The study required patients, whether or not their response to treatment was overtly positive, to build an AI program, which Dr. Berman says captures the nuances encountered in clinical care. said it might not.
For example, a person with cervical dystonia may have pain-dominant symptoms, but if treatment does not improve their most troublesome symptoms, is the patient really benefiting? It may improve symptoms, but it doesn’t result in real improvement in dystonia, he said, adding, “To understand how AI programs like this could enhance our approach to treatment for individual patients. , more research is needed,” he added.
Dr. David Vaillancourt, who led a large international study demonstrating that AI can help identify types of parkinsonism, said such programs would never replace the clinical judgment of neurologists and other physicians. says no.
“This is another tool in our tool kit to help neurologists,” says Dr. Vaillancourt, professor in the Department of Applied Physiology and Exercise at the University of Florida. “You can see things that the eye can’t see and put together a lot of data.”
Allison Brashear, MD, MBA, FAAN, professor of neurology, vice president of health sciences, and dean of the Jacobs School of Medicine and Biomedical Sciences at the University of Buffalo, said he welcomes objective biomarkers of treatment efficacy.
“There were no biomarkers for focal dystonia,” she said.
Disclosure
Dr. Simonyan received consulting fees from AbbVie unrelated to his current research. Doctor. Berman, Brashear, and Mills had no disclosure.