Accurate personalized survival prediction for amyotrophic lateral sclerosis patients

Li-Hao Kuan, Pedram Parnianpour, Rafsanjany Kushol, Neeraj Kumar, Tanushka Anand, Sanjay Kalra, Russell GreinerScientific 2023 · DOI

Abstract

Amyotrophic Lateral Sclerosis (ALS) is a rapidly progressive neurodegenerative disease. Accurately predicting the survival time for ALS patients can help patients and clinicians to plan for future treatment and care. We describe the application of a machine-learned tool that incorporates clinical features and cortical thickness from brain magnetic resonance (MR) images to estimate the time until a composite respiratory failure event for ALS patients, and presents the prediction as individual survival distributions (ISDs). These ISDs provide the probability of survival (none of the respiratory failures) at multiple future time points, for each individual patient. Our learner considers several survival prediction models, and selects the best model to provide predictions. We evaluate our learned model using the mean absolute error margin (MAE-margin), a modified version of mean absolute error that handles data with censored outcomes. We show that our tool can provide helpful information for patients and clinicians in planning future treatment.

Introduction: The Challenge of ALS Prognosis

Amyotrophic Lateral Sclerosis (ALS) is a devastating neurodegenerative disease characterized by a median survival of 2–5 years. However, patient outcomes are notoriously variable, ranging from a few months to over a decade. This high variability makes it nearly impossible for clinicians to provide accurate prognostic counseling using traditional methods.

Most existing survival models provide either a "risk score" (ranking patients by relative risk) or a binary classification (e.g., predicting survival at a fixed 18-month mark). While academically useful, these metrics often fail to provide the practical information needed for clinical planning, such as deciding when to transition to hospice or scheduling interventions like non-invasive ventilation (NIV). Furthermore, survival data is frequently "censored"—where patients leave a study before an event occurs—which can bias standard machine learning models. This research addresses these gaps by predicting the specific timing of "composite respiratory failure" (defined as death, tracheostomy, or significant NIV use) while accounting for censored data.

Methods: The SuperLearner Framework

The researchers developed a SuperLearner, an ensemble-like approach that automatically selects the best-performing configuration from a pool of candidate algorithms. The framework evaluated four base models: Accelerated Failure Time (AFT), Cox Proportional Hazard, Multitask Logistic Regression (MTLR), and Random Survival Forest (RSF).

Individual Survival Distributions (ISDs)

A key innovation of this tool is the output of Individual Survival Distributions (ISDs). Instead of a single risk score or a point estimate, the model generates a probability curve for each patient. This allows clinicians to determine the probability of survival at any future time point. For example, a user could identify the specific month by which a patient has a 90% probability of remaining event-free, facilitating proactive care planning.

Individual Survival Distributions (ISDs) for three distinct patients compared to the population-level Kaplan-Meier estimator.

To evaluate the model's performance on censored data, the authors utilized the MAE-margin, a modified version of Mean Absolute Error. Since the exact event date is unknown for censored patients, the MAE-margin estimates it using the mean of the Kaplan-Meier survival curve, conditioned on the patient's last known follow-up.

Datasets and Cohort Characteristics

The study utilized data from the Canadian ALS Neuroimaging Consortium (CALSNIC), a prospective multicenter study. The final analysis included 172 patients after preprocessing. The feature set comprised 114 clinical variables (such as ALSFRS-R scores, age, and symptom duration) and 68 MRI-based atlas regions for cortical thickness.

Distributions of survival time, censor time, and patient age within the study cohort.

Notably, 45% of the patients in this cohort were censored, highlighting the importance of using evaluation metrics like MAE-margin that can handle incomplete outcome data.

Key Insights and Results

The SuperLearner's internal cross-validation identified Multitask Logistic Regression (MTLR) as the superior base learner. MTLR is particularly effective for ISDs as it treats survival as a series of dependent binary outcomes across multiple time intervals.

Clinical vs. Imaging Features

A surprising finding was that clinical features alone outperformed models that included MRI-based cortical thickness data. While cortical thinning is a known hallmark of ALS, it did not provide statistically significant additional predictive power in this specific dataset, likely due to the relatively small sample size and the high dimensionality of the imaging data.

Performance Metrics

The final model demonstrated significant improvements over baseline population averages:

  • Accuracy: Achieved an MAE-margin of 14.2 months, compared to 17.4 months for the baseline.
  • Discrimination: Achieved a C-index of 0.70, indicating a strong ability to rank patient survival order.
  • Binary Benchmarks: When converted to binary predictions, the model was 77% accurate for 1-year survival and 85% accurate for 3-year survival.

Limitations and Future Directions

While promising, the study notes several limitations. The cohort of 172 patients is small for complex machine learning, which may have limited the utility of MRI data. Additionally, the current model is static, using data only from the initial clinical visit. Future iterations could incorporate longitudinal data to capture disease progression over time. Finally, exploring more granular neuroimaging modalities, such as vertex-wise cortical thickness or Diffusion Tensor Imaging (DTI), may eventually reveal the prognostic value of brain imaging in ALS.

Summary updated

Abstract

Amyotrophic Lateral Sclerosis (ALS) is a rapidly progressive neurodegenerative disease. Accurately predicting the survival time for ALS patients can help patients and clinicians to plan for future treatment and care. We describe the application of a machine-learned tool that incorporates clinical features and cortical thickness from brain magnetic resonance (MR) images to estimate the time until a composite respiratory failure event for ALS patients, and presents the prediction as individual survival distributions (ISDs). These ISDs provide the probability of survival (none of the respiratory failures) at multiple future time points, for each individual patient. Our learner considers several survival prediction models, and selects the best model to provide predictions. We evaluate our learned model using the mean absolute error margin (MAE-margin), a modified version of mean absolute error that handles data with censored outcomes. We show that our tool can provide helpful information for patients and clinicians in planning future treatment.
BibTeX
@inproceedings{kuan2023accurate,
  author = {Li-Hao Kuan and Pedram Parnianpour and Rafsanjany Kushol and Neeraj Kumar and Tanushka Anand and Sanjay Kalra and Russell Greiner},
  title = {{Accurate personalized survival prediction for amyotrophic lateral sclerosis patients}},
  booktitle = {Scientific},
  year = {2023},
  doi = {10.1038/s41598-023-47935-7},
}
Figure