Differential Dementia Diagnosis on Incomplete Data with Latent Trees

Christian Ledig, Sebastian Kaltwang, Antti Tolonen, Juha Koikkalainen, Philip Scheltens, Frederik Barkhof, Hanneke Rhodius-Meester, Betty M. Tijms, Afina W. Lemstra, Wiesje M. van der Flier, Jyrki Lötjönen, Daniel RueckertMICCAI 2016 · DOI

TL;DR

This paper introduces a novel method using Latent Trees (LT) to improve differential dementia diagnosis when patient data is incomplete. By effectively imputing missing features, their approach significantly enhances classification accuracy compared to traditional methods, even with substantial data loss.

What Problem Does This Paper Address?

Accurate diagnosis of neurodegenerative diseases like dementia is crucial for effective treatment and research. However, clinical datasets often suffer from missing information due to various practical reasons, such as missed examinations or measurement errors. Many standard machine learning classifiers, like Support Vector Machines (SVMs), require complete feature sets for both training and testing. While some methods exist to handle missing data, such as removing incomplete samples (reducing statistical power and discarding valuable data) or simple imputation techniques (like using the mean), these approaches can introduce bias or fail to capture complex feature dependencies. This paper tackles the challenge of robustly diagnosing dementia in the presence of incomplete data, aiming to leverage all available information without compromising diagnostic performance.

Methods

The core of this work is the application of Latent Tree (LT) models for feature imputation. Unlike Gaussian Mixture Models (GMMs) which assume features depend on a single latent variable, LT models learn a hierarchical structure of latent variables, allowing them to capture more complex dependencies between features. This hierarchical structure is represented as a tree, where observed features are leaves and latent variables are internal nodes.

The authors adapt an existing LT structure learning algorithm to handle missing data in two key ways:

  1. Training on Incomplete Data: The LT model is trained to estimate the joint distribution of all features, even when some are missing in the training set.
  2. Imputation for Testing: For new, unseen data with missing features, the LT model infers the most likely values for these missing features based on the observed ones and the learned hierarchical structure.

Once the missing features are imputed, a standard discriminative classifier, specifically an SVM, is used for the final diagnosis. The paper also explores using the LT model's likelihood measure to detect outlier samples.

Datasets

The study utilizes a dataset of 491 patients from the Amsterdam Dementia Cohort at the VU University Medical Center. This dataset includes:

  • VIS (VIsual assessment): 13 biomarkers assessed during clinical visits (e.g., age, gender, education, MMSE, ApoE4 genotype, Fazekas Score, presence of lacunes/infarcts, atrophy measures).
  • IMG (IMaging data): 18 features automatically derived from MRI scans (e.g., unnormalized volumes of brain structures like hippocampus, amygdala, entorhinal cortex, ventricles, and vascular burden measures).

The data covers five diagnostic categories: Subjective Cognitive Decline (SCD), Alzheimer's Dementia (AD), Frontotemporal-Lobe Dementia (FTD), Dementia with Lewy Bodies (DLB), and Vascular Dementia (VaD).

Key Insights and Ideas

  • LTs Excel at Capturing Feature Dependencies: The hierarchical structure of LTs allows them to learn complex relationships between features, leading to more accurate imputation of missing values compared to methods that assume simpler dependencies (like GMMs or mean imputation).
  • Robustness to Missing Data is Crucial: The paper demonstrates that even with a significant portion of features missing (up to 50% in both training and testing), the LT-based approach maintains a high level of diagnostic accuracy, highlighting its practical utility.
  • Imputation Enhances Classification: Improved feature imputation directly translates to better performance for downstream classification tasks, underscoring the importance of addressing missing data effectively.
  • Potential for Outlier Detection: The LT model's inherent ability to measure sample likelihood provides a valuable tool for identifying potentially erroneous or inconsistent data points within a dataset.

Key Results

  • Improved Feature Imputation: LT imputation showed significantly lower Normalized Root Mean Squared Error (NRMSE) compared to mean and GMM imputation across selected features, especially when a substantial percentage of features were missing. For example, with 20% missing features, LT achieved an NRMSE of 2.0 for MMSE, compared to 2.9 for mean and 2.8 for GMM.
  • Enhanced Classification Accuracy: When 50% of features were missing in both training and testing data, LT+SVM achieved a balanced accuracy of 56.6%, significantly outperforming Mean+SVM (53.7%) and GMM+SVM (52.6%). This accuracy is notably higher than the random guess level (20% for 5 classes).
  • Outlier Detection Performance: The LT model demonstrated strong performance in detecting artificially introduced erroneous samples (via feature swapping), achieving a high AUC of ~87% for outlier detection, outperforming GMM (~83%).

Limitations and Future Work

  • Limitations:
    • The paper notes that the advantage of LT imputation decreases with increasing percentages of missing data, as there is less information available to leverage the learned structure.
  • While LT outperforms other methods, the classification accuracy still drops significantly with extensive missing data, indicating that there's an upper bound to what imputation can achieve.
  • The authors acknowledge that feature swapping can sometimes result in samples that appear valid (e.g., swapping left and right hippocampal volumes), which can limit the perfect detection of all outliers.
  • Future Work:
    • The paper suggests that the LT model's generative nature could be further exploited for the classification of arbitrary, a-priori unknown targets.
  • The open-source implementation of LT is made available, encouraging further research and application of this method.
  • Further exploration of the LT model's potential in other domains with incomplete data is implied.

Summary updated

BibTeX
@inproceedings{ledig2016differential,
  author = {Christian Ledig and Sebastian Kaltwang and Antti Tolonen and Juha Koikkalainen and Philip Scheltens and Frederik Barkhof and Hanneke Rhodius-Meester and Betty M. Tijms and Afina W. Lemstra and Wiesje M. van der Flier and Jyrki Lötjönen and Daniel Rueckert},
  title = {{Differential Dementia Diagnosis on Incomplete Data with Latent Trees}},
  booktitle = {International Conference on Medical Image Computing and Computer-Assisted Intervention},
  year = {2016},
  doi = {10.1007/978-3-319-46723-8\_6},
}
Figure