Applying latent class cluster analysis & data mining methods to identify classes of CFS patients...predictive of treatment success, 2022, Clapperton

Looks like one can download the PhD thesis for free immediately. But it is probably quite technical. I wonder does it use the PACE trial data? Goldsmith was involved with that as well as Chalder.

Applying latent class cluster analysis and data mining methods to identify classes of chronic fatigue syndrome patients that are predictive of treatment success

https://ethos.bl.uk/OrderDetails.do?uin=uk.bl.ethos.863375

Title: Applying latent class cluster analysis and data mining methods to identify classes of chronic fatigue syndrome patients that are predictive of treatment success

Author: Clapperton, Ben

Awarding Body: King's College London

Current Institution: King's College London (University of London)

Date of Award: 2022

Availability of Full Text:

Access from EThOS:
Immediate download. Please login to continue.

Access from Institution:
Link to institutional repository

Abstract:

Chronic fatigue syndrome (CFS) is a condition characterised by chronic disabling fatigue and other symptoms that currently cannot be explained by any underlying medical condition.

It is estimated that 250,000 of the UK population suffer from CFS.

Although clinical trials support the effectiveness of cognitive behaviour therapy in terms of fatigue and physical functioning, the success rate for individual patients is modest.

Patients vary in their response and little is known about which factors predict or moderate their treatment outcomes.

Identifying moderators of treatment effect is typically done in a regression-based approach by assessing interactions between clinical and other baseline variables and treatment groups.

Moderators typically are of small effect size and prediction models often show poor accuracy in predicting outcomes.

My project took a different approach to identify classes of patients with similar baseline characteristics.

I compared a model-based cluster analysis method of Latent Class Analysis against an automatic distribution-free machine learning algorithm, Self-Organising Maps.

The classes identified were tested for predictive usefulness of treatment effects.

Characteristics of the classes can be used to inform clinicians about the types of individuals who benefit from specific treatments.

The suitability of the clustering methods was compared using data on CFS patients in two datasets, a large clinical cohort study and a randomised clinical trial.

Using the trial data, I also compared the performance of the clustering techniques against a computer-intensive statistical learning penalised regression method which provides predictions without clustering.

The LASSO regression model was also developed to identify moderators of treatment success. These comparisons allow the assessment of the potential advantages of clustering approaches and their capability of identifying complex relationships between variables in the prediction of treatment success in CFS above the regression-based approach.

Supervisor: Stahl, Daniel Richard ; Chalder, Trudie ; Goldsmith, Kimberley Ann Sponsor: Not available

Qualification Name: Thesis (Ph.D.) Qualification Level: Doctoral

EThOS ID: uk.bl.ethos.863375 DOI: Not available

 

"clinical trials support the effectiveness of cognitive behaviour therapy in terms of fatigue and physical functioning"

!!!!!

 

LMFAO at the "machine learning" using hundreds of data points. This is basically like trying to build a house using 4 planks, half a nail and gum stuck on a shoe. Pathetic.

Machine learning has only started showing useful results by building on billions of data points. And it requires objective and precise data points, it doesn't work on wishy-washy stuff. You may as well feed the horoscopes of 30 people to this and it would be just as useless.

Also you can't start from a false premise. CBT obviously doesn't "work", no matter how meaningless a definition of work is. This is a very good example of misusing machine learning, of simply not understanding the conditions that make it work, which are still very restricted. ML does not work with unsolved problems, you need what's called a "ground truth" to compare to to build the error function.

If you take, for example, AlphaFold, the system that solved protein folding, it worked by training on known proteins that have already been solved. It doesn't work without having known solutions to the problem.

In fact, this is basically a very sad example of what Adam Douglas meant about 42 being the answer to life, the universe and everything: it doesn't matter how smart your machine is, if you ask the wrong question the answer will be useless.

Might as well put some quantum this and that in here. Garbage pseudoscience.

 

Very strange abstract. Describes the methodology but not the results.

 

Ok I just skimmed through the method and discussion. None of the ML techniques identified useful clusters or predictors of treatment response.

 

The author concludes, more or less, that the clusters identified from the rich and detailed psychological data end up being useless because they yielded clusters of disparate sizes, with no apparent correlation to treatment outcomes.

His “final thoughts” below are telling:


Prediction performance in all of the models I fitted was clinically not useful and poor with little moderating effect. Precision medicine has made progress in many medical areas, but research in complex diseases such as Chronic Fatigue Syndrome has not yet benefited from the new methodological advances. In recent years, these advances have formed an integral part of other medical areas, especially in oncology. More research is needed in this area with a focus to identify data and theory-driven types that are coherent across symptoms, brain regions, genes, physiology and behaviour, and are relevant to a clinical outcome such as response to treatment (Stahl & Stamate, 2018).

Anyone with a dogmatic belief in the value of cluster analysis might therefore view this thesis as severely undermining the legitimacy of the PACE data and the battery of psychological tests conducted on patients. If the data cannot tell us anything useful, even under severe interrogation, they would conclude, it must be meaningless.

It would be amusing if it attracted citations in that context. Chalder would have been better advised to have had it suppressed.

 

They’ll probably never allow him to publish it and unpublished theses are rarely if ever cited.

 

The supervisors know that the vast majority of ppl only read abstracts. So if the abstract omits key info or is misleading, you can essentially bury bad news.

 

Ah. Yes. And while it is available on Kings’ servers, it could presumably be withdrawn at any time. Perhaps someone with a relevant blog could stick it on google docs and publish a link, along with a salient summary? Technically a breach of copyright, but hardly actionable.

 

Obviously you can't do precision medicine using BS questionnaires, which have zero precision and aren't even objective values. Same as you can't do ML using a miniscule amount of data that consists of biased arbitrary questions. Big data isn't some buzzword, the data sets used to train useful systems are enormous.

The only response from a ML undergraduate to this project should have been: this doesn't work for this, doesn't even apply to this kind of problem.