Machine Learning Identifies Three ALS Progression Patterns

Researchers at the Tours ALS center in France have developed a machine learning framework that identifies three distinct progression patterns in patients with amyotrophic lateral sclerosis. The study, published in BioData Mining, utilized a weakly supervised learning approach to analyze longitudinal functional data from 353 patients. By learning a custom distance metric for clustering, the team aimed to stratify patients based on their disease trajectories, offering a potential tool for earlier prognostic assessment and personalized care planning.

Amyotrophic lateral sclerosis is a progressive neurodegenerative disease characterized by the degeneration of motor neurons, with a median survival of three to five years following symptom onset. Current treatments, such as Riluzole, offer limited benefits in slowing progression, and the clinical variability of the disease makes prognosis difficult. Patients differ significantly in their site of onset, progression rate, and cognitive involvement, which complicates management and often delays critical interventions like non-invasive ventilation or gastrostomy. The identification of reliable biomarkers and clustering methods is therefore crucial for improving both diagnosis and long-term care strategies.

The study analyzed data from the Tours ALS center registry, covering patients monitored between April 1997 and October 2023. After rigorous preprocessing to handle missing or aberrant data, 353 patients who were monitored every three months between January 2004 and July 2023 were retained for the analysis. These patients provided ALS Functional Rating Scale-Revised scores, along with clinical, biological, and demographic data. The exclusion criteria included patients with fewer than five assessments, those without subscores, and those with clinically improbable score increases, ensuring the reliability of the longitudinal sequences used for clustering.

The core of the methodology involved a weakly supervised learning approach to define a distance between patient pairs based on their ALSFRS-R sequences. Instead of relying on predefined thresholds or single time-point scores, the researchers extracted descriptive variables from the temporal sequences, including monitoring duration, initial scores, and overall slopes. A sigmoid function was also fitted to estimate specific values, such as the score at one year and the D50 value, which represents the time to halve functional ability. These features were used to train a classifier to learn a weighted distance measure, referred to as the Weak-Supervised Distance, which captures the relative importance of different sequence characteristics.

Using this learned distance, the team applied various clustering algorithms, including K-means, K-medoids, and Agglomerative Hierarchical Clustering, with and without dimensionality reduction via Uniform Manifold Approximation and Projection. The workflow WSD_UMAP_KMD_3, which combined UMAP with K-medoids to form three clusters, emerged as the optimal configuration. This workflow achieved a silhouette score of approximately 0.6, indicating strong separation between clusters, and demonstrated high stability with an Adjusted Rand Index of 0.7. The three clusters corresponded to slow, moderate, and fast progression patterns, with the slow progression cluster showing a median survival of 32 to 38 months, compared to 23 to 27 months for the fast progression cluster.

Survival analysis using the Kaplan-Meier estimator confirmed that the identified clusters had distinct prognostic values. Over 50% of patients in the slow progression cluster survived beyond 50 months, whereas less than 25% of patients in the other two clusters reached this milestone. The log-rank test revealed significant differences in survival among the clusters, with p-values well below the 0.05 threshold. This separation suggests that the clustering method effectively captures clinically meaningful subpopulations, allowing for a more nuanced understanding of disease heterogeneity than previously possible with standard threshold-based approaches.

The researchers also investigated which diagnostic parameters could predict cluster membership at the time of diagnosis. A Random Forest model was trained on 13 significant parameters, including ALSFRS-R subscores, IgG levels, albumin quotient, and time to diagnosis. The model achieved a specificity and negative predictive value of approximately 0.75, with sensitivity and positive predictive value close to 0.5. While these results are promising, they indicate that early prediction of progression patterns remains challenging due to the inherent variability in initial clinical presentations and the limited availability of biological markers at diagnosis.

Comparisons with state-of-the-art approaches, including those by Gomeni, Grollemund, and Meyer, showed that the proposed method offered superior cluster balance and separation. Previous methods often relied on fixed thresholds for ALSFRS-R scores, which can be sensitive to minor variations and may not reflect meaningful clinical differences. The new framework’s ability to integrate multiple sequence characteristics provided a more robust classification, although the predictive performance was comparable to some existing methods. The study highlights the trade-off between the number of clusters and the balance of patient distribution, with three clusters offering the best compromise between interpretability and statistical validity.

Limitations of the study include the single-center nature of the dataset and the reduction in cohort size due to strict preprocessing criteria. The absence of external validation means that the generalizability of the findings requires further confirmation in larger, multi-center cohorts. Additionally, the exploratory nature of the workflow selection may introduce optimistic bias in performance estimates. The researchers acknowledge that the predictive model was trained on a restricted subset of variables due to missing data, which may limit the identification of other potential predictors. Future work should focus on external validation and the integration of additional biomarkers, such as neurofilament light chain, to enhance the model’s accuracy and clinical utility.

Despite these limitations, the study provides a proof-of-concept for using machine learning to stratify ALS patients based on longitudinal functional data. The identified clusters align with known progression patterns, offering a framework for earlier patient stratification in clinical trials and personalized care. By leveraging routinely collected data, this approach could help clinicians anticipate functional decline and plan interventions more proactively, ultimately improving outcomes for patients with this complex and heterogeneous disease.

To establish patient similarities, the researchers derived specific descriptors from the functional rating sequences before applying machine learning algorithms. The longitudinal monitoring recorded visits at an average interval of 87 days with a standard deviation of 66 days. Beyond monitoring duration and baseline scores, the sequence analysis tracked both overall decline and the steepest drop observed between consecutive clinic visits. For unaligned timeline visits, the researchers fitted a sigmoid function to interpolate longitudinal progression, specifically calculating the functional score at twelve months, the percentage decline at six months, and the time required to reach a score of twenty-four, defined as the D50 value marking a halving of functional ability. The non-linear least squares sigmoid estimation demonstrated high fidelity across the cohort, achieving a median error of 1.43 points on the functional scale, which corresponds to an approximate deviation of three percent. A probabilistic voting system aggregated expert labeling functions into training pairs for a linear support vector machine, producing the customized distance metric.

The data curation process began with an initial cohort of 913 patients registered at the Tours facility, which was systematically narrowed down through strict quality controls. The investigators excluded 375 individuals who had fewer than five clinical consultations, 12 patients lacking recorded functional subscores, 154 individuals whose trajectories showed clinically implausible score increases exceeding two points between consecutive evaluations, and 19 patients whose functional monitoring began more than 30 days after initial consultation. The remaining 353 patients yielded 62,128 pairwise comparisons. To prevent statistical distortion, the authors filtered diagnostic parameters, eliminating 99 clinical variables and 145 longitudinal biological parameters that presented more than 50 percent missing entries. They also applied distinct correlation cutoffs, setting a 0.7 Spearman threshold to eliminate redundant pairwise trajectory variables and a 0.9 threshold for diagnostic clinical and laboratory measurements.

The final clustering partition separated the 353 participants into groups of 139, 121, and 93 individuals, achieving an entropy score of 0.987 that reflected balanced patient distribution across categories. In addition to distinct trajectory slopes, the subgroups revealed distinct clinical tendencies at initial presentation. Patients assigned to the slow progression group entered monitoring with initial scores consistently near the scale ceiling of 48. In contrast, baseline scores in the moderate and fast decline groups exhibited wider dispersion. Although general demographic indicators such as age at onset, sex, and weight showed no statistically significant differences across clusters, anatomical presentation varied. Bulbar-onset cases were predominantly located within the rapid decline cluster, spinal-onset cases concentrated heavily within the intermediate cluster, and no patients with primary respiratory onset were assigned to the slow deterioration group.

Feature importance analyses conducted with the Random Forest classifier clarified the relative utility of baseline clinical and laboratory measurements. The rate of functional progression before diagnosis demonstrated the highest predictive weight, closely followed by the total diagnostic delay and forced vital capacity measurements, whereas the bed turning subscore provided the lowest predictive influence. Laboratory testing identified only two biological markers with statistically significant associations across clusters: the cerebrospinal fluid immunoglobulin G concentration and the albumin quotient. Biochemically, the albumin quotient quantifies the permeability of the blood-brain and blood-cerebrospinal fluid barriers, while cerebrospinal fluid immunoglobulin levels reflect localized immune system reactivity. Their retention in the predictive model suggests that central nervous system inflammation and barrier disruption correlate with distinct clinical progression paths, even though overall biological parameters demonstrated lower predictive weight than physical metrics.

Subject of Research: Machine learning clustering of ALS patient progression trajectories

Article Title: Learning a distance for the clustering of patients with amyotrophic lateral sclerosis

Article References: Dominguez, G. T. Y., Alarcan, H., Peralta, V., Labroche, N., Corcia, P., Marcel, P., & Blasco, H. (2026). Learning a distance for the clustering of patients with amyotrophic lateral sclerosis. BioData Mining, 19(1), Article 78. https://doi.org/10.1186/s13040-026-00579-5

Image Credits: AI Generated

DOI: 10.1186/s13040-026-00579-5

Keywords: amyotrophic lateral sclerosis, machine learning, patient clustering, ALSFRS-R, prognosis, neurodegenerative disease, weakly supervised learning, clinical trial design, Learning, distance, clustering, patients