Source-linked AI summary
Disentangling Representation using Attributes-based Gaussian Estimation for Medical Sound Diagnosis
Ke Zhao
TL;DR
Deep models can lack interpretability and exhibit bias across demographic and other attribute differences, complicating medical diagnosis. AGEDR maps attributes into VAE latents and disentangles attribute from non-attribute information using mutual-information minimization. It outperformed conventional classification and disentangled-representation comparators in the reported COUGHVID experiments.
Problem
Deep learning’s automatically extracted features can lack interpretability, and demographic, lifestyle, medical-history, and regional differences can affect performance and applicability.
Source: §I · p. 1Method
AGEDR maps scalar attributes to Gaussian latent representations, aligns them with part of a VAE latent vector, and disentangles that part from remaining latents using mutual-information minimization.
Source: §I · p. 2, §II · p. 2, A. Attributes Mapping Embedding · p. 2Results
AGEDR outperformed conventional classification models and similar disentangled-representation methods; the concatenated-latent group achieved 92.52% accuracy.
Source: C. Performance Comparision · p. 4Takeaways & Limitations
The experiments provide evidence that AGEDR reduces attribute-related information to mitigate attribute-related domain shifts and reduce unnecessary sensitive information.
Source: §I · p. 2
Abstract
from arXiv · showhide
Deep learning has a powerful capability of feature extraction. However, the lack of fairness and interpretability in deep neural networks poses limitations to their adoption in the medical domain. This paper proposes a disentangled representation learning (DisenRL) framework, named the Attributes-based Gaussian Estimation for Disentangled Representation (AGEDR), which incorporates Attribute Mapping Embedding (AME) modules designed to map attributes into vectors and align them with a subset of the latent vectors in a Variational AutoEncoder (VAE). This part of the latent vector will be disentangled from the remaining latent vectors by minimizing mutual information. A classifier is then trained using the mean parameters of the latent vectors from the VAE. Extensive experiments demonstrate that AGEDR outperforms both conventional classification models and existing disentangled representation learning methods. The ablation experiments also indicate the disentangling capability and fairness of AGEDR. The source code is publicly available at https://github.com/ZhaoKe1024/DisentangledRepr.
I. INTRODUCTION
The introduction frames interpretability, bias, and domain shift as challenges for deep learning in medical diagnosis. It motivates AGEDR as a way to map attributes into VAE latents and disentangle attribute-related information.
I. INTRODUCTION
- Deep models automatically extract features but lack interpretability, while demographic, lifestyle, medical-history, and regional differences can affect performance and applicability.The introduction identifies these biases and domain shift as practical challenges.Source: §I · p. 1
- Disentangled representation learning seeks decomposable latent vectors to address bias and domain shift, while supervised methods can align vectors with actual semantics.The paper presents this alignment as addressing the lack of interpretability.Source: §I · p. 1
- Prior approaches concatenate attributes (InfoGAN), train attribute predictors [2], impose equivariance, or use normalizing flows and information-theoretic methods.The introduction notes that these approaches rely on prior attributes or structures.Source: §I · p. 1
- Task-trained attribute-related latents lack a prior distribution for probabilistic density estimation, limiting direct calculation of DKL and mutual information.The paper proposes AME to map scalar attributes into normal parameters, then uses VAE latents zα and zβ and mutual-information minimization to disentangle them.Source: §I · pp. 1–2
- AGEDR combines attribute-to-Gaussian mapping and mutual-information disentanglement, with experiments on COUGHVID examining classification, attribute-information reduction, and fairness.The stated contributions include two ablation experiments and empirical evidence related to attribute-related domain shifts and sensitive information.Source: §I · p. 2
II. PROPOSED METHOD
The proposed method uses a VAE with normally distributed latent vectors and information-theoretic alignment and disentanglement. Its AME module maps scalar attributes to Gaussian latent representations, which are partitioned from non-attribute information.
II. PROPOSED METHOD
- The method uses a VAE whose latent vectors follow N(0, 1), with information-theoretic approaches for semantic alignment and disentangled representation.The framework trains the VAE through reconstruction and later uses encoder mean parameters for disease prediction.Source: §II · p. 2, Figure 2 · p. 3
A. Attributes Mapping Embedding
- AME maps scalar attributes into richer vectors, then estimates Gaussian parameters and samples attribute latent vectors through reparameterization.For categorical attributes, one-hot values select a column from a learned matrix; the resulting vector has dimension n(i).Source: A. Attributes Mapping Embedding · p. 2
- The VAE latent vector is split into attribute latents zα, formed by concatenating the r attribute vectors, and non-attribute latents zβ for remaining sample information.The non-attribute portion has dimension n(0).Source: A. Attributes Mapping Embedding · p. 2
- The method aligns the AME and VAE attribute representations by minimizing DKL, enabling the VAE to extract attribute information in zα without additional attribute annotation.The DKL term is included as a regularizer in the loss.Source: A. Attributes Mapping Embedding · p. 2, Figure 2 · p. 3
B. Disentangled Latent Vector
Because inputs contain information about all attributes, the attribute and non-attribute latent vectors overlap; the method uses mutual-information minimization to reduce this overlap, approximating MI with an evidence lower bound because it is difficult to compute directly.
B. Disentangled Latent Vector
- Overlapping information between attribute latent vector zα and non-attribute latent vector zβ is linked to reduced robustness and interpretability.The inputs encompass information about all attributes, so the two latent vectors remain entangled before disentanglement.Source: B. Disentangled Latent Vector · p. 3
- Mutual-information minimization is used as the disentanglement loss to decrease information overlap between zα and zβ.Because MI is difficult to calculate directly and attribute-vector lengths vary, the method optimizes an evidence lower bound and transforms the loss into Ldisen.Source: B. Disentangled Latent Vector · p. 3
- The loss terms use encoder distributions for zα and zβ across input samples, with q(x|z) representing reconstruction from a latent vector.The first and second terms derive from the MI definition and Jensen’s inequality; the third term comes from the MI ELBO.Source: B. Disentangled Latent Vector · p. 3
C. The Classification Task
The classifier uses the VAE encoder’s mean parameters rather than random latent samples, and the classification loss uses focal loss.
C. The Classification Task
- The classifier takes the encoder’s mean parameter μ as input because noise-induced randomness in VAE latent vectors negatively affects prediction tasks.The passage contrasts this prediction impact with the benefit of latent randomness for generation.Source: C. The Classification Task · p. 3
- The classification loss Lcls uses focal loss for the classification task.The final model combines four loss terms with weights w1–w4 and three additional weights v1–v3.Source: C. The Classification Task · p. 3
- The final model framework is depicted in Figure 2.The passage identifies the four-term loss and its weights as part of the final model.Source: C. The Classification Task · p. 3
A. Dataset
The study uses filtered and segmented COUGHVID cough audio, retaining expert-annotated data and deriving health and COVID-19 samples for classification.
A. Dataset
- COUGHVID contains 34,434 cough audios with crowdsourced user annotations and more than 2,800 samples with expert annotations.The authors note that the data were not collected professionally and include unreliable samples.Source: A. Dataset · pp. 3–4
- Filtering for complete expert annotations, a status field, cough-detection probability of at least 0.8, and good quality retained 720 samples.These criteria were applied to address unreliable samples in the dataset.Source: A. Dataset · p. 4
- The 720 samples were segmented at mute clips, cough segments under 0.36 seconds were discarded, and the resulting 2,850 segments were standardized to 1.46 seconds at 22,050 Hz.Longer segments were center-cropped and shorter ones zero-padded.Source: A. Dataset · p. 4
- The final set contains 2,076 health and 774 COVID-19 samples, represented as 64-by-128 log-Mel spectrograms with cough-type and severity labels.Cough type has three categories and severity has four.Source: A. Dataset · p. 4
B. Model Settings
The model uses convolutional VAE modules with attribute-specific latent dimensions for cough type and severity, plus a fully connected classifier.
B. Model Settings
- The convolutional VAE has a 30-dimensional latent vector, with 6 dimensions for cough type, 8 for severity, and 16 non-attribute dimensions.The classifier is a two-layer fully connected network by default.Source: B. Model Settings · p. 4
- The loss weights are w1=2.0, w2=0.3, w3=0.0025, w4=0.01 and v1=1.0, v2=1.25, v3=5.0; module learning rates are 0.0003, 0.0001, and 0.0002 for AME, VAE, and classifier, respectively.All modules use Adam.Source: B. Model Settings · p. 4
- Training ran for 370 epochs until model parameters tended to converge, with focal loss used to address the roughly 2.68:1 health-to-COVID-19 class imbalance.The passage refers to Figure 3 for the values of each loss-function component.Source: B. Model Settings · p. 4
C. Performance Comparision
AGEDR is reported to outperform conventional and DisenRL baselines, with 92.52% accuracy and 99% recall. Tested classifier heads performed similarly, with SVM sometimes higher.
C. Performance Comparision
- The VAE, disentanglement, attribute, and classification loss components all show decreasing trends during iteration.Source: Figure 3 · p. 4
- A two-layer neural network and SVM classification head achieved similar performance, with SVM sometimes higher.The evaluation compares MobileNetV2, AST, DisenIB, and IDB-SR as baselines.Source: C. Performance Comparision · p. 4
- 92.52% accuracy and 99% recall were achieved by AGEDR using concatenated attribute and non-attribute latent vectors.The paper also reports higher performance than similar DisenRL methods.Source: Table I · p. 4, C. Performance Comparision · p. 4
D. Disentanglement Study
The disentanglement study reports reduced classification performance when the attribute latent vector is omitted. Its t-SNE visualization shows class-label clustering but no discernible clustering by the displayed attributes.
D. Disentanglement Study
- Without zα, precision was 91.45% and recall was 75.87%.The authors interpret the decline as evidence that AME carries attribute-related information and that Ldisen disentangles zα.Source: D. Disentanglement Study · p. 4, Table II · p. 5
- The t-SNE visualization shows disease classification corresponding to class labels, while cough type and severity annotations are distributed without discernible clustering.The authors interpret this pattern as an attribute-desensitization effect on classification.Source: Figure 4 · p. 5, D. Disentanglement Study · p. 5
E. Alignment Effect
AME-mapped attributes achieved 99.21% clustering accuracy for both cough type and severity. The conclusion describes AGEDR as disentangling latent attributes while maintaining classification accuracy, and proposes deep Gaussian mixture models for future work.
E. Alignment Effect
- 99.21% clustering accuracy was achieved for both cough type and severity attributes using k-means on latent vectors.The alignment experiment maps attributes into the implicit vector space and aligns them with part of the VAE latent vector using KL divergence.Source: E. Alignment Effect · p. 5
IV. CONCLUSION
- AGEDR is described as disentangling latent attributes while maintaining classification accuracy, with interpretability, fairness, and reliability as its medical-audio-diagnosis significance.Source: §IV · p. 5
- Future work proposes using a deep Gaussian mixture model to handle unstructured data and mine higher-order combination features.The stated aim is stronger disentanglement capability and diagnostic performance on more complex data.Source: §IV · p. 5