Optimised machine learning for time-to-event prediction in healthcare applied to timing of gastrostomy in ALS : a multi-centre, retrospective model development and validation study
(2025) In EBioMedicine 121.- Abstract
Background: Amyotrophic lateral sclerosis (ALS) is invariably fatal but there are large variations in the rate of progression. The lack of predictability can make it difficult to plan clinical interventions. This includes the requirement for gastrostomy where early or late placement can adversely impact quality of life and survival. Methods: We designed a model to predict the timing of gastrostomy requirement in ALS as indicated by 5% weight loss from diagnosis. We considered >5000 different prediction model configurations including spline models and a set of deep learning (DL) models designed for time-to-event prediction. The optimal prediction model was chosen via a Bayesian framework to avoid overfitting. Model covariates were... (More)
Background: Amyotrophic lateral sclerosis (ALS) is invariably fatal but there are large variations in the rate of progression. The lack of predictability can make it difficult to plan clinical interventions. This includes the requirement for gastrostomy where early or late placement can adversely impact quality of life and survival. Methods: We designed a model to predict the timing of gastrostomy requirement in ALS as indicated by 5% weight loss from diagnosis. We considered >5000 different prediction model configurations including spline models and a set of deep learning (DL) models designed for time-to-event prediction. The optimal prediction model was chosen via a Bayesian framework to avoid overfitting. Model covariates were measurements routinely collected at diagnosis; a separate longitudinal model also incorporated weight at six months. We employed a training dataset of 3000 patients from Europe, and two external validation cohorts spanning distinct populations and clinical contexts (United States, n = 299; and Sweden, n = 215). Missing data was imputed using a random forest model. Findings: The optimal model configuration was a logistic hazard DL model. The optimal model achieved a median absolute error (MAE) between predicted and measured time of 3.7 months, with AUROC 0.75 for gastrostomy requirement at 12 months. To increase accuracy we updated predictions for those who had not received gastrostomy at six months after diagnosis: here MAE was 2.6 months (AUROC 0.86). Combining both models achieved MAE of 1.2 months for the modal group of patients. Prediction performance is stable across both validation cohorts. Missing data was imputed without degrading model performance. Interpretation: To enter routine clinical practice a prospective study will be required, but we have demonstrated stable performance across multiple populations and clinical contexts suggesting that our prediction model can be used to guide individualised gastrostomy decision making for patients with ALS. Funding: Research Ireland (RI) and Biogen have supported the PRECISION ALS programme.
(Less)
- author
- Weinreich, Marcel
; McDonough, Harry
; Heverin, Mark
; Domhnaill, Éanna Mac
; Yacovzada, Nancy
; Magen, Iddo
; Cohen, Yahel
; Harvey, Calum
; Elazzab, Ahmed
; Gornall, Sarah
; Boddy, Sarah
; Alix, James J.P.
; Kurz, Julian M.
; Kenna, Kevin P.
; Zhang, Sai
; Iacoangeli, Alfredo
; Al-Khleifat, Ahmad
; Snyder, Michael P.
; Hobson, Esther
; Chio, Adriano
; Malaspina, Andrea
; Hermann, Andreas
; Ingre, Caroline
; Costa, Juan Vazquez
; van den Berg, Leonard
; Panadés, Monica Povedano
; van Damme, Philip
; Corcia, Phillipe
; de Carvalho, Mamede
; Al-Chalabi, Ammar
; Hornstein, Eran
; Elhaik, Eran
LU
; Shaw, Pamela J.
; Hardiman, Orla
; McDermott, Christopher
and Cooper-Knock, Johnathan
(Less) - organization
- publishing date
- 2025-11
- type
- Contribution to journal
- publication status
- published
- subject
- keywords
- Amyotrophic lateral sclerosis (ALS), Gastrostomy, Machine learning, Personalised medicine, Time-to-event prediction
- in
- EBioMedicine
- volume
- 121
- article number
- 105962
- publisher
- Elsevier
- external identifiers
-
- pmid:41075354
- scopus:105018102861
- ISSN
- 2352-3964
- DOI
- 10.1016/j.ebiom.2025.105962
- language
- English
- LU publication?
- yes
- id
- 35a4a52e-580c-4d72-93a1-a84542d989a2
- date added to LUP
- 2025-11-24 13:49:13
- date last changed
- 2026-01-06 02:15:04
@article{35a4a52e-580c-4d72-93a1-a84542d989a2,
abstract = {{<p>Background: Amyotrophic lateral sclerosis (ALS) is invariably fatal but there are large variations in the rate of progression. The lack of predictability can make it difficult to plan clinical interventions. This includes the requirement for gastrostomy where early or late placement can adversely impact quality of life and survival. Methods: We designed a model to predict the timing of gastrostomy requirement in ALS as indicated by 5% weight loss from diagnosis. We considered >5000 different prediction model configurations including spline models and a set of deep learning (DL) models designed for time-to-event prediction. The optimal prediction model was chosen via a Bayesian framework to avoid overfitting. Model covariates were measurements routinely collected at diagnosis; a separate longitudinal model also incorporated weight at six months. We employed a training dataset of 3000 patients from Europe, and two external validation cohorts spanning distinct populations and clinical contexts (United States, n = 299; and Sweden, n = 215). Missing data was imputed using a random forest model. Findings: The optimal model configuration was a logistic hazard DL model. The optimal model achieved a median absolute error (MAE) between predicted and measured time of 3.7 months, with AUROC 0.75 for gastrostomy requirement at 12 months. To increase accuracy we updated predictions for those who had not received gastrostomy at six months after diagnosis: here MAE was 2.6 months (AUROC 0.86). Combining both models achieved MAE of 1.2 months for the modal group of patients. Prediction performance is stable across both validation cohorts. Missing data was imputed without degrading model performance. Interpretation: To enter routine clinical practice a prospective study will be required, but we have demonstrated stable performance across multiple populations and clinical contexts suggesting that our prediction model can be used to guide individualised gastrostomy decision making for patients with ALS. Funding: Research Ireland (RI) and Biogen have supported the PRECISION ALS programme.</p>}},
author = {{Weinreich, Marcel and McDonough, Harry and Heverin, Mark and Domhnaill, Éanna Mac and Yacovzada, Nancy and Magen, Iddo and Cohen, Yahel and Harvey, Calum and Elazzab, Ahmed and Gornall, Sarah and Boddy, Sarah and Alix, James J.P. and Kurz, Julian M. and Kenna, Kevin P. and Zhang, Sai and Iacoangeli, Alfredo and Al-Khleifat, Ahmad and Snyder, Michael P. and Hobson, Esther and Chio, Adriano and Malaspina, Andrea and Hermann, Andreas and Ingre, Caroline and Costa, Juan Vazquez and van den Berg, Leonard and Panadés, Monica Povedano and van Damme, Philip and Corcia, Phillipe and de Carvalho, Mamede and Al-Chalabi, Ammar and Hornstein, Eran and Elhaik, Eran and Shaw, Pamela J. and Hardiman, Orla and McDermott, Christopher and Cooper-Knock, Johnathan}},
issn = {{2352-3964}},
keywords = {{Amyotrophic lateral sclerosis (ALS); Gastrostomy; Machine learning; Personalised medicine; Time-to-event prediction}},
language = {{eng}},
publisher = {{Elsevier}},
series = {{EBioMedicine}},
title = {{Optimised machine learning for time-to-event prediction in healthcare applied to timing of gastrostomy in ALS : a multi-centre, retrospective model development and validation study}},
url = {{http://dx.doi.org/10.1016/j.ebiom.2025.105962}},
doi = {{10.1016/j.ebiom.2025.105962}},
volume = {{121}},
year = {{2025}},
}