Computer-Assisted Levothyroxine Dose Selection for the Treatment of Postoperative Hypothyroidism

[helvella]

A paper which causes me significant concern.

This is my blog which explains, I hope, my deep scepticism about existing prediction of required levothyroxine doses.

helvella - Estimation of Levothyroxine Dosing in Adults

A discussion about the use of formulas to estimate levothyroxine dosing. Includes link to a downloadable spreadsheet which calculates several of these.

helvella.blogspot.com/p/hel...

But at least the existing approaches have an understandable formula. And, despite me producing a spreadsheet, there is nothing that cannot easily be worked out with at most a calculator, pen and paper.

Maybe it is simply because the publisher has erected a paywall, but we do not have access to the formulas behind this approach. We cannot even start to examine the computer-assisted dose selection. Nor the formulas. Nor the documentation surrounding it. Nor the theoretical basis.

That its goal is defined in terms of TSH is clearly not supported by much theory and many papers.

Their choice of factors is questionable.

Their approach to validation needs to be justified. If even one patient receives a worse outcome by following their algorithm, that patient ends up in a difficult position. Could they demand some form of compensation? And were they fully informed before allowing their health to be controlled by an obscure algorithm?

I am reminded of some of the issues with Body Mass Index. BMI is often quoted in terms of weight and height. But it needs far more thought than that. Applying basic BMI calculations to a one-legged person immediately shows up its naivety.

Computer-Assisted Levothyroxine Dose Selection for the Treatment of Postoperative Hypothyroidism

Martin Barrio, Christopher D. Raeburn, Robert McIntyre, Jr., Maria Albuja-Cruz, Bryan R. Haugen, and Nikita Pozdeyev

Published Online: 20 Apr 2023 doi.org/10.1089/thy.2023.0033

Abstract

Background: Thyroid hormone replacement with levothyroxine (LT4) is a recommended treatment for patients undergoing thyroidectomy. The starting LT4 dose is frequently calculated based on the patient's weight. However, the weight-based LT4 dosing performs poorly in clinical practice, with only ∼30% of patients achieving target thyrotropin (TSH) levels at the first thyroid function testing after treatment initiation. A better way to calculate the LT4 dose for patients with postoperative hypothyroidism is needed.

Methods: In this retrospective cohort study we used demographic, clinical, and laboratory data for 951 patients after thyroidectomy and several regression and classification machine learning methods to develop an LT4 dose calculator for treating postoperative hypothyroidism targeting the desired TSH level. We compared the accuracy with the current standard-of-care practice and other published algorithms and evaluated generalizability with fivefold cross-validation and out-of-sample testing.

Results: The retrospective clinical chart review showed that only 285/951 (30%) patients met their postoperative TSH goal. Obese patients were overtreated with LT4. An ordinary least squares regression based on weight, height, age, sex, calcium supplementation, and height:sex interaction predicted prescribed LT4 dose in 43.5% of all patients and 45.3% of patients with normal postoperative TSH (0.45-4.5 mIU/L). The ordinal logistic regression, artificial neural networks regression/classification, and random forest methods achieved comparable performance. LT4 calculator recommended lower LT4 doses to obese patients.

Conclusions: The standard-of-care LT4 dosing does not achieve the target TSH in most thyroidectomy patients. Computer-assisted LT4 dose calculation performs better by considering multiple relevant patient characteristics and providing personalized and equitable care to patients with postoperative hypothyroidism. Prospective validation of LT4 calculator performance in patients with various TSH goals is needed.

Keywords: levothyroxine; machine learning; postoperative hypothyroidism.

liebertpub.com/doi/full/10....

[diogenes Remembering]

Yet again another TSH-must-be-in range fallacious paper. Also, one cannot accurately predict the rate at which the tissues react to oral dosing of T4 or indeed how effective T4 use is. A proposition built on sand is the best I can give this.

[tattybogle]

My daughter just discovered a button on her company car .. push it and it parallel parks itself !

it does it very well , presumably it's controlled by AI that's been taught it's 'Target' is:

"close to, but not on the pavement , not on top of any other cars, or cats or children"

AI that thinks it's Target is "get TSH in range" would park a lot of people in a ditch .

[helvella Administrator Thyroid UK in reply to tattybogle]

I've not used automated parallel parking.

But I do wonder how it would cope with:

Deep water running down gutter;

Snow;

Extremely slippery road surface such as ice;

A flat tyre.

All circumstances which human drivers are expected to cope with or avoid.

[tattybogle in reply to helvella]

dunno ..... i will report back in winter (assuming the human doesn't bend it first)

if they are going to research an AI method to look for shorter time to correct dose (based on TSH ) , then i think it would be good science to research it against a 'control' ....

for example against some AI who's target is "shorter time to no remaining thyroid symptoms"..... Now that WOULD be interesting .. if they did that then AI might be able to teach the endo's something useful.

[jimh111]

The objective is fine, guidance for estimating an initial dose for people who have their thyroid removed. By inputting half a dozen measurements into a computer program they can improve their accuracy from 30 to 43.5 percent! A lot of bother to achieve a small increase in a first stab at establishing a therapeutic dose. Doctors won't have time to mess about with this. It doesn't even matter if the initial dose is a little out, patients take time to settle down.

There are two basic errors. TSH is a poor indicator of thyroid status (except for initial diagnosis) and monotherapy is a bad idea. My view is levothyroxine monotherapy increases cardiac and cancer risks and for this reason should not be routinely used

[helvella Administrator Thyroid UK in reply to jimh111]

No - afraid I flat out disagree.

Yes - I very much agree TSH is the wrong target and everything associated with that. And I think I agree about T4 monotherapy.

But the very example I put in my post highlights an issue. Weight/height stats for a one-legged person are substantially different to those of someone with the full set.

If their AI system does NOT have the possibility of inputting special features such as limb loss, and nothing in the abstract suggests their brains have noticed this, then someone dumbly tonking in the patient's details will get a potentially seriously wrong answer.

A real doctor looking at a real patient might twig the issue. But when you start down the AI route, you have to include all these things right from the start. Otherwise the AI will "learn" either without any one-legged patients being recognised and just applying standard data to them, or using the data from one-legged patients to inform treatment of all patients.

No - I have no idea how many hypothyroid patients are one-legged. It is just a fairly obvious issue that could easily be very badly mishandled. And there will be many more such issues which need to be considered.

[jimh111 in reply to helvella]

What I'm saying is that trying to use a complex system is silly (and pseudo science). May as well rely on a simple system based on weight, or starting people on 100 mcg. It's just to get an initial dose to start people off. Not targeting a TSH level, just avoiding unnecessary suffering from the old idea of starting low and titrating up.

As someone who studied mathematics I get very annoyed by the misuse of mathematics, using it to give a false impression of precision. These formulae are not AI. AI used intelligently can be useful in medicine. An example is the use of AI in spotting cancer in scans. Trying to use AI to replace proper in person consultations would be another example of abuse of mathematics / computing.

This study is a dreadful waste of resources, our only consolation is that anyone daft enough to spend time on such work isn't much use to patients.

[arTistapple]

Just when you think things couldn’t possibly get any worse!

[DippyDame]

When will these wretched people stop treating human beings like machines....

[Zephyrbear in reply to DippyDame]

The only machines patients are in some parts of the world, and probably even in the UK, are cash machines… keep ‘em sick, big pharma benefits!

[DippyDame in reply to Zephyrbear]

That too!