Parkinson’s Disease (PD) is diagnosed by a cluster of symptoms related to its effect on motor nerves. There is no lab test or imaging study that makes the diagnosis. As a result, it often takes some time for a vague symptom to become so well established that a physician can identify it. Consequentially, the ability to make an early diagnosis is difficult. And because the disease is characterized by its symptoms, it is difficult for a physician to track PD’s progress. After all, can you objectively say there is more weakness or a greater tremor based on subjective physical examinations months apart?
The current study used an artificial intelligence methodology to analyze the breathing patterns during sleep of roughly 7,600 patients and predicted the presence or development of PD in 90% of cases. But before jumping into the particulars of the study, a bit of background on PD.
Parkinson’s Disease – a bit of context
“So slight and nearly imperceptible are the first inroads of this malady, and so extremely slow its progress, that it rarely happens, that the patient can form any recollection of the precise period of its commencement. The first symptoms perceived are, a slight sense of weakness, with a proneness to trembling in some particular part; sometimes in the head, but most commonly in one of the hands and arms.
… In this stage, the sleep becomes much disturbed. The tremulous motion of the limbs occur during sleep, and augment until they awaken the patient, and frequently with much agitation and alarm.”
– An Essay on the Shaking Palsy James Parkinson, Member of the Royal College of Surgeons 1817
Parkinson’s Disease (PD) is a degenerative disease of our neurologic system that results in tremors and bradykinesia. Brady, for slow, kinesia, for movement, reflects an increasing inability to move our muscles; over time, it results in the rigidity of motion and alterations in our posture. One way you might think of the muscular changes is the fine tremors, often first seen in the hands or feet. In place of our normal smooth motion, there is just a slight fluctuation in position resulting in a slight variation in hand position – the tremor. Metaphorically, the shock absorbers that smooth out our ride are not working at their peak, and our ride becomes slightly bumpier.
Respiratory alterations are also part of PD, although they are not as noticeable early in the disease and may be overlooked. To extend the analogy of the hand tremors, perhaps early in the course of PD, our normally smooth inhalation and exhalation, both governed by muscles, may evidence tiny fluctuations, a staccato alteration that we do not notice. While the underlying cause of PD remains unknown, it appears that it involves loss of dopaminergic activity in the brainstem, which controls, among other functions, our respiration.
Some patients with PD describe the sensation of dyspnea (difficulty in breathing), and shortness of breath – a symptom we believe is based upon the perceived strength of our respiratory muscles. In PD, there appears to be a blunted response to our levels of CO2 and O2, both of which play a role in the regulation of our breath. The data are unclear as to which of these gases may be impacted in individual cases of PD. Still, it does seem that these alterations result in early dyscoordination of our breath – what I have described as a staccato alteration in the usually smooth flow of air with inhalation and exhalation.
“PD is frequently associated with respiratory disturbances, even in pre-motor stages, … the presence of breathing symptoms should alert the physician of a PD not well controlled or in progression.”
Researchers used a “breathing belt,” basically a gauge measuring the change in size, placed around the chest or abdomen of a sleeping patient recording the movement while they slept.  They used a combination of datasets from sleep studies aggregating those 7,600 patients, totaling 12,000 nights of sleep or 120,000 hours of sleepy-time. Approximately 10% of those individuals carried a diagnosis of PD.
As with most AI research, the dataset, consisting of the sleep signals from just one night of testing, was divided between a training group and a test group. And as with AI, what was identified in those signals is unknown. That said, the “black box” produced some significant results.
- It accurately identified nearly 90% of PD patients. It identified almost 80% of patients without PD (the diagnostic sensitivity) and an 80% specificity, correctly identifying patients with the disease. For comparison, digital mammography, widely used to screen and identify breast cancer, has a sensitivity of 85% and a specificity of 90%. Keep in mind that we are talking about PD, where until now, the diagnosis rested on a physician’s experience and examination skills.
- The current means of tracking the progression of PD is the MDS-Unified Parkinson’s Disease Rating Scale (MDS-UPDRS), a quasi-subjective multipart evaluation. The AI analysis of those nocturnal breathing signals was highly correlated, 0.94 (1.0 is identical), with an individual’s MDS-UPDRS score. So the algorithm may also track progression.
- Using a much smaller dataset of individuals undergoing sleep tests who had no evidence of PD on initial testing but had a diagnosis of PD on the subsequent examination six years later, the algorithm correctly identified 75% of those initially asymptomatic patients who went on to PD. This suggests that the algorithm can be used in early detection.
- Finally, one of the weaknesses of AI-derived diagnostics is that they are frequently not as reliable when tested against other datasets than those used during training and initial testing. This was not the case here, where the diagnostic ability of the algorithm, when tested against completely different sleep data, showed a diagnostic accuracy close to 90%
There is, of course, at least one caveat (isn’t there always?). These respiratory changes are not specific or definitive (pathognomic is the medical term), solely for PD but may be seen in other motor neuron neurologic disorders. That said, this can represent a significant advance in diagnosis and care. It is non-invasive, it could be done at home, and it identifies and tracks patients in an objective way that will augment, if not supplant, our current subjective measures.
“no physician today can detect PD or assess its severity from breathing. This shows that AI can provide new clinical insights that otherwise may be inaccessible.”
For a disease that is not currently curable, where we can improve the quality of life with medications but continue to stumble in making the initial diagnosis and following the course of the disease, this piece of AI is a step forward.
 They could also perform the same measurements using radio waves without needing any breathing belt or other wearable.
Sources: Respiratory dysfunction in Parkinson’s disease: a narrative review European Respiratory Journal Open Research DOI: 10.1183/23120541.00165-2020
Artificial intelligence-enabled detection and assessment of Parkinson’s disease using nocturnal breathing signals Nature Medicine DOI: 10.1038/s41591-022-01932-x