Tag Archives: AI

Will PSP research benefit from AI? It already is.

Posted on by

Some of you may be wondering why my blogging has fallen off lately, with only two posts in July, two in August and none so far halfway through September. Besides enjoying my summer, I’ve been writing other things:


• Two invited editorials to accompany journal publications – one on a new brain MRI technique for diagnosing PSP, the other on a set of physical exam findings with good diagnostic performance in the early years of PSP;
• A 63-slide PowerPoint lecture on the atypical Parkinsonian disorders (APDs) that I delivered last week;
• A chapter on the same topic for the 14th edition of a neurology textbook;
• A chapter on PSP for the 4th edition of a textbook on movement disorders; and
• A share of a collaborative paper on imaging, blood and skin biopsy diagnostics for the APDs.
For each task, I used the AI apps ChatGPT, Gemini and Claude to help organize my thinking and to quickly find things in the literature I may have missed. Of course, I had to double-check the results, but it was still a huge timesaver.


A couple of months ago, one of this blog’s faithful readers (Jack Phillips, Chairman of the Board of CurePSP) asked me if AI is being used in the fight against PSP. I told him how I’m using it as a writing tool, and more important, that it’s helping the search for new drugs and new diagnostic tests. Here’s a more detailed answer:


• AI apps can be taught to recognize all the parts of the brain on MRI. If they’re told which are from people with PSP and which are not, they can figure out which brain areas make that differentiation most effectively. They can then apply those insights to MRIs from future patients, particularly those whose diagnoses remain uncertain to their doctors. The type of MRI abnormality most useful here is focal atrophy – shrinkage of specific, damaged areas. But other types of changes such as scarring and iron deposition can also be recognized.


• The ten different sub-types of PSP have subtle differences in which brain areas are affected most on MRI. Using the same reasoning as for the PSP vs non-PSP task mentioned above, AI can then provide a good guess as to which sub-type is at work. This is important early in the disease course because most clinical treatment trials are confined to the PSP-Richardson syndrome subtype, which accounts for only half of all PSP. Also, the different subtypes develop differently over the years and have different survival durations – information useful to clinicians in counseling patients and families.


• The disease process of PSP spreads through the brain not uniformly like water through a dry sponge, but through routes determined by synaptic connections and kinds of contacts. What’s actually spreading is the tau protein in mis-folded form. Positron emission tomographic (PET) imaging can show the precise locations where tau is most concentrated. That can be coupled with MRI using a new AI machine-learning technique called Subtype and Stage Inference (SuStaIn). Imagine having only one tau PET scan and one MRI scan from each of hundreds of patients, each with known dates of symptom onset and of the scans. Then you have to figure out the time course and routes of the disease spread. The SuStain algorithm is told the dates of symptom onset and of the scans, measures the severity of the PET and MRI abnormalities in each of dozens of brain areas and puts all that temporal and spatial data together to create a kind of three-dimensional “movie” of the spread of the disease over time. That could allow finer assessment of differences between PSP sub-types, provide a new outcome measure for neuroprotection trials, and provide clues as to what makes some brain areas more resistant than others to the disease process.


• SuStain and similar highly sensitive measures of PSP progression could provide a much more sensitive measure of benefit of potential disease-slowing drugs. In this way, a trial could require far fewer patients and far shorter time spans than at present. I can envision a future where a trial using SuStain could require only a dozen patients and 6 months. If the drug shows a subtle slowing of the disease relative to placebo, the molecular structure of the drug could be tweaked (also with the help of AI) and another 6-month round of testing could start — a far cry from the 4 years it takes to test one drug, a combination of drugs could be evaluated, and components of the cocktail could be dropped and/or added for the next round.


• I happen to know from painful experience that extracting information from one patient’s medical records for the purpose of guiding subsequent clinical care is a major chore and doing it for dozens of patients in a research trial is worse. It’s child’s play, however, for an AI-based indexing algorithm, even when the records are in different formats or handwritten. This is especially relevant for rare diseases like PSP, where patients in a trial are likely to have been referred from multiple physicians from different health systems using different record formats.


• In 1990, the first high-resolution image of a single protein molecule was produced using cryogenic electron microscope (cryo-EM). Here’s an image of a mis-folded tau protein molecule from someone with PSP (from Shi et al. Nature 2021).

Each little bump Is an amino acid. As far as we know, the folding pattern is the same in every tau molecule in every brain cell in every part of the brain in every patient with PSP. It’s a very different folding pattern for CBD despite its frequent outward resemblance to PSP. Each little nook and cranny is a potential spot for a drug to attach to prevent this toxic form of tau from interacting with other molecules, aggregating with other, identically folded tau molecules, or templating its abnormality onto normal copies of tau. Whichever, mechanism is chose, the disease could theoretically be halted in its tracks. If given the order of amino acids in the tau molecule (which is well known), and the amino acids at each little nook and cranny (also well known), AI could design a molecule to fit. It would be a monkey wrench in the PSP works.

One thing is for sure – this graph is not going to trend down any time soon:

Artificial intelligence and artificial CSF drainage

Posted on by

A common diagnostic problem is distinguishing PSP from normal-pressure hydrocephalus (NPH), but a new way to look at brain MRIs using artificial intelligence could have the solution.  I’ll now administer the usual large dose of background information:

NPH occurs in the same age group as PSP but is much more common.  Its three classic features sound a lot like PSP: frontal cognitive loss, urinary incontinence, and gait impairment.  But those often don’t appear until late in the course and other issues such as general slowness, reduced vertical eye movement and tremor can precede them.  Of course, those also are shared with PSP.

In NPH, the fluid-filled cavities of the brain enlarge and over-stretch the brain’s fibers to produce the symptoms.  The cause of the cerebrospinal fluid (CSF) accumulation in many cases is a partial blockage of its normal absorption into the blood.  In some cases, that appears to be the result of scarring from an old episode of brain infection or bleeding around the brain.  Other cases, called “idiopathic NPH,” have no history of such inflammatory events and their cause remains unknown.  There is also in NPH some evidence of a neurodegenerative component, as in PSP, Parkinson’s, and Alzheimer’s.

A diagnosis of NPH depends less on the clinical history and exam than on two other things: 1) a specific pattern on MRI of brain tissue loss and enlarged CSF spaces and 2) benefit after removal of some CSF.  I’ll discuss those in turn:

A. MRI diagnosis. Below are MRI images from idiopathic NPH (middle row) and PSP (bottom row).  The main differences between PSP and NPH are indicated by the labels on the right.  PSP features widening of the spaces between the brain’s folds caused by atrophy of the brain tissue.  But in NPH, the spaces toward the top of the brain are as tight as, or tighter than, normal.  There are other, less reliable, MRI differences, none of them adequately sensitive or specific for NPH.

B. CSF diagnosis. The other diagnostic feature is the response to CSF drainage.  It’s not just a diagnostic test; it also predicts the likely response to treatment by shunting.  If someone in whom NPH is suspected has an MRI consistent with NPH and no signs of other potential causes of their symptoms, the physician will usually perform a spinal tap to remove about 30-50 ml with before-and-after videos of the gait and other actions. (The average adult has about 100-150 ml at any one time, but the daily turnover is about 500 cc, so the 30-50 ml loss is replaced after only a few hours.)  Some neurologists prefer the greater diagnostic reliability provided by a more prolonged period of drainage via a soft plastic tube temporarily inserted into the lumbar CSF space (the same place where the needle of a spinal tap goes), but this can have complications.

Whichever temporary method of CSF removal is used, a good symptomatic response would prompt consideration of a tube, called a “shunt,” permanently implanted into the brain to direct flow of some CSF, usually into the abdominal cavity.  Of course, implanting such a shunt into the brain can produce complications such as infection or bleeding, so we’d first like to make sure the person doesn’t have PSP, which offers no potential shunt benefit to compensate for that risk.

I should point out that PSP is far from the only disease that can mimic NPH and not respond to shunting.  Among the others are the far more common PD and AD.  That means that only a small minority of “NPH candidates” actually has NPH, so placing brain shunts in all the candidates would be highly inadvisable, to put it mildly. So, it’s important to make the right diagnosis.

Over the decades since 1965, when NPH was first described in the literature, the number of proposed diagnostic methods has been prodigious and none has been sufficiently accurate.  But now, the cavalry may have arrived in the form of AI.  A group of researchers led by Drs. Fubuki Sawa and Syoji Kobashi of the University of Hyongo in Japan has used “convolutional neural networks,” a form of deep learning, to produce a predictive model. It used the most specific and informative MRI features from 59 people with NPH who subsequently benefitted from shunting and 65 people with PSP by current, validated criteria.  The resulting statistical formula produced a perfect score of 1.000 in the area under the curve (AUC) of the receiver operating characteristic (ROC). (Wikipedia has a nice little explanation of that statistic here.  Basically, it’s the ability of a diagnostic test to minimize both false positives and false negatives, with 1.0 being perfect and 0.5 being equivalent to a coin toss. Its virtue is that it’s applied to an individual, not merely to the averages of two groups.) 

Perhaps easier to intuit is the test’s accuracy, according to Sawa et al, of 0.983.  That statistic is formally defined as the fraction of all the participants who received a correct diagnosis from the formula. Such power for a diagnostic test is nearly unheard-of in medicine, but keep in mind that the definition of NPH in this study wasn’t autopsy, but an MRI showing the typical features plus a response to CSF shunting.  So that means that the input and outcome variables were partly redundant, inflating the accuracy to some extent.

The other caveat is that this technique only distinguished PSP from NPH, not from PD or anything else.  But the general AI-based statistical technique should be applicable to many kinds of diagnostic situations where the two candidate diseases cause atrophy in different parts of the brain.  We eagerly await those papers from Drs. Sawa and Kobashi, and we hope, others.

The take-home if you’re someone with PSP:

  1. Should people with a diagnosis of PSP get a new MRI each year in the hope that a pattern of NPH will emerge and a shunt procedure confer improvement?  Probably not, because an MRI showing the abnormalities of PSP won’t change into the abnormalities of NPH over time.   
  2. Should people with PSP get a shunting procedure just in case they actually have NPH?  Definitely not, as the risk of both short- and long-term shunt complications far exceeds the likelihood of benefit.
  3. Instead of either of these: Keep hydrated and well-nourished, avoid falls and aspiration, minimize unnecessary medications with your doctor’s advice and consent, get some exercise, maintain a social life, and join an FDA-approved clinical trial if one is available.
  4. Also, consider getting a second diagnostic opinion from a neurologist subspecializing in movement disorders, who can scrutinize the original MRI for evidence of NPH that might have eluded the original neurologist or radiologist.

    The take-home if you’re a neurologist:

    At each follow-up visit or phone call, keep in mind the possibility that the diagnosis may not actually be PSP, but something much more treatable — like NPH. Then, work up or refer accordingly.

    Sing, o muse!

    Posted on by

    You’ll recall from my 7/4/24 post that Rep. Jennifer Wexton of Virginia announced last year that she has PSP.  Below is a new press release from her office.  My point relates to its final paragraph.

    Wexton Marks Disability Pride Month on House Floor Using New AI Voice Model

    Washington, July 25, 2024

    Washington, DC — Today, Congresswoman Jennifer Wexton (D-VA) delivered a speech on the floor of the House of Representatives to recognize Disability Pride Month, celebrated each July. Wexton utilized her new AI voice model for the first time on the House floor to share how she has fought throughout her career for disability rights and accessibility issues, and how that has now become a personal fight for her since her Progressive Supranuclear Palsy (PSP) diagnosis last year.

    “I hope that when people see me continuing to do all I can to keep living my life and doing this job I love as best I can, they understand and appreciate the courage, resilience, and spirit that so many Americans of differing abilities demonstrate each and every day. I hope I can be a voice — even an AI voice — for Americans facing accessibility challenges and other disabilities, because too often people only see us for that disability. And in truth we are so much more,” said Wexton.

    The AI voice model was created through ElevenLabs, a software company that specializes in developing natural-sounding speech synthesis software using deep learning. The model was generated consensually using a collection of previous public speeches from Rep. Wexton provided by her team. Any additional questions about the ElevenLabs technology can be directed to press@elevenlabs.io.

    More than the nominal topic of Disability Pride Month, the text mentions that she used an AI-based speech aid to deliver a speech on the floor of the House.  It used audio recordings of her public speeches before PSP’s dysarthria set in to train the AI software to translate a written text into speech in her own voice, not the electronic, synthetic voice we all recognize as “computer speech.”  This is important because the distinctive timbre of one’s speech provides an important emotional identifier. 

    This technique, relying as it does on a prepared, written text, is not suitable for ordinary two-way conversations unless you can type extremely fast. Few people with PSP can do that, even with the help of word-completion software.  Besides, hardly anyone has a collection of old audio clips on which to “train” the AI software that Rep. Wexton used.  An audio bank deliberately created for this purpose in an early stage of illness is one potential solution.

    I wanted to review the latest on this topic, mostly to help me advise CurePSP about what sort of programs or research it might support going forward.  So, I consulted my newest muse, ChatGPT.  Here’s what it had to say about voice banking and other solutions for severe dysarthria:

    You can bank your own voice for future use in a speech generation app in case you become disabled. Voice banking involves recording a variety of phrases and sentences while your voice is still healthy. These recordings can later be used to create a synthetic version of your voice, allowing speech generation apps to produce speech that sounds like you.

    Here are the general steps to bank your voice:

        Choose a Voice Banking Service: Select a service that offers voice banking. Some popular options include VocaliD, Acapela, and ModelTalker.

        Record Phrases: Follow the service’s instructions to record a set of phrases and sentences. The more you record, the better the quality of the synthesized voice.

        Submit Recordings: Upload your recordings to the voice banking service.

        Voice Synthesis: The service will process your recordings and create a synthetic version of your voice.

        Integration with Speech Generation App: Once your synthetic voice is ready, it can be integrated into a speech generation app. This allows you to communicate using your own voice even if you lose the ability to speak naturally.

    Voice banking is a proactive step that can significantly improve the quality of life for individuals facing conditions that may affect their ability to speak.

    Alternatives to voice banking?

    If voice banking is not a suitable option for you, there are several alternatives that can help with communication if you become disabled. Here are a few options:

    1. Voice Cloning

    Voice cloning uses advanced AI to create a synthetic version of your voice with fewer recordings. Services like Respeecher and Descript can create high-quality voice models from a limited amount of audio data.

    2. Text-to-Speech (TTS) with Customization

    Text-to-speech technology can be customized with different voices, accents, and tones. Some TTS systems allow you to adjust parameters to make the voice sound closer to your own. Popular TTS engines include:

        Amazon Polly

        Google Text-to-Speech

        Microsoft Azure Speech Service

    3. Pre-Recorded Messages

    You can record common phrases and sentences that you might need. These recordings can be played back as needed. Devices like the GoTalk series or AAC (Augmentative and Alternative Communication) devices often support this feature.

    4. Alternative and Augmentative Communication (AAC) Devices

    AAC devices and apps provide various ways to communicate through symbols, text, and synthetic speech. Examples include:

        Tobii Dynavox

        Proloquo2Go

        TouchChat

    5. Predictive Text Apps

    Predictive text apps can assist in quicker communication by predicting the words and phrases you are likely to use. These apps learn from your typing habits to offer more accurate suggestions. Notable apps include:

        SwiftKey

        Gboard

    6. Communication Boards and Books

    Low-tech options such as communication boards or books with pictures, symbols, or letters can be very effective, especially in environments where technology might not be available.

    7. Eye-Tracking Technology

    For individuals with severe physical disabilities, eye-tracking technology allows communication through eye movements. Devices like the Tobii Eye Tracker can be used to control computers and AAC software.

    8. Sign Language and Gestures

    If you retain some physical ability, learning sign language or using a system of personalized gestures can be an effective way to communicate.

    Each of these alternatives offers different advantages depending on your specific needs and circumstances. Exploring these options can help you find the most effective means of communication in the event of disability.

    Now back to human me. 

    It goes without saying that these solutions have to be tailored to the individual – no easy task considering that the disease itself worsens with time.  Another complication is that many of the techniques require manual dexterity or eye movement control beyond the abilities of many people with PSP.  (It also goes without saying that anything from ChatGPT can be incomplete or even wrong.)

    Medicare pays most of the cost of some of these things if there’s a doctor’s prescription. Discuss with a speech pathologist, which you should be doing anyway for general PSP-related speech and swallowing issues.

    If you consider me your PSP muse, whose mine? Now I know.