AdaMedGraph: Adaboosting Graph Neural Networks for Personalized Medicine
Extracted source text · arXiv 2311.14304
% use the "wcp" class option for workshop and conference
% proceedings
%\documentclass[gray]{jmlr} % test grayscale version
%\documentclass[tablecaption=bottom]{jmlr}% journal article
\documentclass[pmlr,twocolumn,10pt]{jmlr} % W&CP article
% \usepackage{geometry}
% \geometry{margins=0.1in,textwidth=7in}
% The following packages will be automatically loaded:
% amsmath, amssymb, natbib, graphicx, url, algorithm2e
%\usepackage{rotating}% for sideways figures and tables
%\usepackage{longtable}% for long tables
% Please add the following required packages to your document preamble:
\usepackage{booktabs}
\usepackage[normalem]{ulem}
\useunder{\uline}{\ul}{}
% The booktabs package is used by this sample document
% (it provides \toprule, \midrule and \bottomrule).
% Remove the next line if you don't require it.
\usepackage{booktabs}
% The siunitx package is used by this sample document
% to align numbers in a column by their decimal point.
% Remove the next line if you don't require it.
\usepackage{siunitx}
\usepackage{dsfont}
\usepackage{xcolor}
% The following command is just for this sample document:
\newcommand{\cs}[1]{\texttt{\char`\\#1}}% remove this in your real article
% The following is to recognise equal contribution for authorship
\newcommand{\equal}[1]{{\hypersetup{linkcolor=black}\thanks{#1}}}
\newcommand{\ours}{\textit{AdaMedGraph}}
% Define an unnumbered theorem just for this sample document for
% illustrative purposes:
\theorembodyfont{\upshape}
\theoremheaderfont{\scshape}
\theorempostheader{:}
\theoremsep{\newline}
\newtheorem*{note}{Note}
% change the arguments, as appropriate, in the following:
\jmlrvolume{LEAVE UNSET}
\jmlryear{2023}
\jmlrsubmitted{LEAVE UNSET}
\jmlrpublished{LEAVE UNSET}
\jmlrworkshop{Machine Learning for Health (ML4H) 2023} % W&CP title
\newcommand{\xufang}[1]{\textcolor{orange}{\textbf{[Xufang: }#1\textbf{]}}}
% The optional argument of \title is used in the header
% \title[AdaMedGraph]{AdaMedGraph: Adaptive Boosted Graph Neural Networks for Personalized Medicine}
\title[AdaMedGraph]{AdaMedGraph: Adaboosting Graph Neural Networks for Personalized Medicine}
% Anything in the title that should appear in the main title but
% not in the article's header or the volume's table of
% contents should be placed inside \titletag{}
%\title{Title of the Article\titletag{\thanks{Some footnote}}}
% Use \Name{Author Name} to specify the name.
% If the surname contains spaces, enclose the surname
% in braces, e.g.
%\Name{John {Smith Jones}} similarly
% if the name has a "von" part, e.g \Name{Jane {de Winter}}.
% If the first letter in the forenames is a diacritic
% enclose the diacritic in braces, e.g. \Name{{\'E}louise Smith}
% \thanks must come after \Name{...} not inside the argument for
% example \Name{John Smith}\nametag{\thanks{A note}} NOT \Name{John
% Smith\thanks{A note}}
% Anything in the name that should appear in the title but not in the
% article's header or footer or in the volume's
% table of contents should be placed inside \nametag{}
% Two authors with the same address
% \author{%
% \Name{Author Name1\nametag{\thanks{A note}}} \Email{abc@sample.com}\and
% \Name{Author Name2} \Email{xyz@sample.com}\\
% \addr Address
% }
% Three or more authors with the same address:
% \author{%
% \Name{Author Name1} \Email{an1@sample.com}\\
% \Name{Author Name2} \Email{an2@sample.com}\\
% \Name{Author Name3} \Email{an3@sample.com}\\
% \Name{Author Name4} \Email{an4@sample.com}\\
% \Name{Author Name5} \Email{an5@sample.com}\\
% \Name{Author Name6} \Email{an6@sample.com}\\
% \Name{Author Name7} \Email{an7@sample.com}\\
% \Name{Author Name8} \Email{an8@sample.com}\\
% \Name{Author Name9} \Email{an9@sample.com}\\
% \Name{Author Name10} \Email{an10@sample.com}\\
% \Name{Author Name11} \Email{an11@sample.com}\\
% \Name{Author Name12} \Email{an12@sample.com}\\
% \Name{Author Name13} \Email{an13@sample.com}\\
% \Name{Author Name14} \Email{an14@sample.com}\\
% \addr Address
% }
% Authors with different addresses and equal first authors:
\author{%
\Name{Jie Lian}\thanks{Work done during an internship in Microsoft Research Asia.}\Email{jlian@connect.hku.hk}\\
\addr The University of Hong Kong, Hong Kong SAR, China
\AND
\Name{Xufang Luo}\thanks{Corresponding author.}\Email{xufluo@microsoft.com}\\
\addr Microsoft Research, China
\AND
\Name{Caihua Shan} \Email{caihua.shan@microsoft.com}\\
\addr Microsoft Research, China
\AND
\Name{Dongqi Han} \Email{dongqihan@microsoft.com}\\
\addr Microsoft Research, China
\AND
\Name{Varut Vardhanabhuti} \Email{varv@hku.hk}\\
\addr The University of Hong Kong, Hong Kong SAR, China
\AND
\Name{Dongsheng Li} \Email{dongsli@microsoft.com}\\
\addr Microsoft Research, China
}
\begin{document}
\maketitle
\begin{abstract}
Precision medicine tailored to individual patients has gained significant attention in recent times. Machine learning techniques are now employed to process personalized data from various sources, including images, genetics, and assessments. These techniques have demonstrated good outcomes in many clinical prediction tasks. Notably, the approach of constructing graphs by linking similar patients and then applying graph neural networks (GNNs) stands out, because related information from analogous patients are aggregated and considered for prediction.
However, selecting the appropriate edge feature to define patient similarity and construct the graph is challenging, given that each patient is depicted by high-dimensional features from diverse sources. Previous studies rely on human expertise to select the edge feature, which is neither scalable nor efficient in pinpointing crucial edge features for complex diseases.
In this paper, we propose a novel algorithm named \ours, which can automatically select important features to construct multiple patient similarity graphs, and train GNNs based on these graphs as weak learners in adaptive boosting. \ours{} is evaluated on two real-world medical scenarios and shows superiors performance.
% Recent releases and publicly available healthcare data collection have yielded extensive datasets encompassing imaging, genetics, clinical assessments, and so on. Leveraging those multi-modality data for accurate disease prediction is important. In this context, graph-based methods have gained prominence by modeling relationships among patients, allowing us to represent individuals as nodes and capture pairwise similarities through edges. However, existing approaches often rely on manual edge selection and focus on a single relationship type.
% In response, we propose an automated graph construction method that considers multiple relations, offering transparency and interpretability for clinical insights. We evaluated our approach in two medical scenarios, and our AdaMedGraph showed superior performance. Our study demonstrates the potential of graph-based models to enhance disease prediction in diverse healthcare applications.
\end{abstract}
\begin{keywords}
Graph neural networks, adaptive boosting, multi-modality, disease progression modelling, disease prediction
\end{keywords}
\section{Introduction}
\label{sec:intro}
Precision (personalized) medicine aims at providing treatments that are best for a specific patient based on individual variations in genes, environmental factors, and lifestyles. It has aroused much attention due to its potential to refine healthcare decisions for each patient. But the goal of precision medicine is hard to realize because of the vast heterogeneity in patient profiles. Thus, it is necessary to consider multi-modal data from different sources for each patient. Concurrently, machine learning (ML) techniques have demonstrated efficacy in analyzing multi-modal data \citep{tu2023towards,zhang2023biomedgpt}. Also, a growing number of large-scale, shared clinical datasets of images, genetics, and assessments \citep{hulsen2019big,shilo2020axes} accelerates the application of ML to personalized clinical prediction tasks, such as predicting disease progression or occurrence \citep{mhasawade2021machine}.
Among various ML methods, graph neural networks (GNNs) are well-suited for considering relationships between individuals, further facilitating personalized modeling and prediction. Specifically, given a dataset containing the data of many patients and each patient having multi-modal data, we can construct a graph with patients as nodes and connect similar patients, where the similarity is determined by the features of the patients. For example, patients of similar ages or same genders can be connected to form a graph \citep{parisot2017spectral}. Then, by utilizing features from the multi-modal data as node features, the GNN can be trained for predictions. This approach ensures that the modeling and prediction for a given patient are informed not only by their individual data but also by data from analogous patients. Prior studies \citep{xing2019dynamic,zheng2022multi} have underscored the efficacy of such models for population analysis and multi-modal data integration in medical fields.
However, selecting the appropriate edge features to define patient similarity is challenging since each patient has high-dimensional features from multi-modal data sources. It is also crucial because selected edge features greatly influence prediction results. Previous works rely on human expertise and prior knowledge to determine edge features. This approach lacks scalability because one needs to find new edge features for a new problem. Moreover, identifying edge features in prediction tasks related to complex diseases can be non-trivial, even for human experts.
To address the above issue, we propose a novel algorithm named \ours, which can automatically select important features to construct multiple patient similarity graphs. Generally, in \ours, GNNs as weak classifiers are iteratively trained in the adaptive boosting (AdaBoost) process. In each round, the most important feature with a certain criteria is selected as the edge feature, and an edge is established when the gap of this feature between two patients is less than a threshold, which is also determined by the algorithm. Then, a GNN is trained based on the constructed graph. Finally, all trained GNNs are ensembled for prediction. Notably, \ours{} can also be compatible with human prior knowledge by involving graphs built by human experts in the final ensemble model. Therefore, inter-individual information and intra-individual features are well-unified in one model, with human efforts on graph building largely relieved by \ours.
% In summary, \textit{AdaMedGraph} can automatically construct patient similarity graphs
% automatic, transparency
We conduct extensive experiments on two real-world medical scenarios, i.e., Parkinson’s disease (PD) progression and metabolic syndrome prediction. \ours{} shows superior performance on almost all tasks compared with some strong baselines.
% The availability of medical data has grown significantly in recent years, as large-scale collaborative initiatives have collected and shared datasets of imaging, genetic, and clinical assessment data \citep{hulsen2019big,shilo2020axes}. In order to harness the wealth of healthcare data available, there is a pressing need for models capable of seamlessly integrating and effectively utilizing data from various modalities. These models must demonstrate enhanced accuracy in predicting disease occurrence and progression, ultimately contributing to the advancement of precision medicine \citep{mhasawade2021machine}.
% In clinical practice, patients frequently demonstrate similarities in their clinical symptoms during certain stages, which often results in their classification under specific disease subsets. However, it is essential to acknowledge that this classification may overlook the considerable diversity in etiology and disease progression that can occur among individual patients over the entire course of the disease. In light of this, we posit that there is a compelling rationale for considering inter-individual and intra-individual data to construct personalized models, which could have substantial clinical significance \citep{zhang2021graph}. Graph methods have been commonly used to construct relationships among patients and improve the classification accuracy of individuals in the medical field, exploring multi-dimensional data and learning patient similarities from various data sources \citep{jiang2020hi, kazi2019inceptiongcn}. Within this particular context, the representation of each subject or patient is as a node, and the modeling of pairwise similarities is accomplished by the establishment of edges that connect these nodes. Prior studies \citep{xing2019dynamic,zheng2022multi} have indicated that these models offer effective means for population analysis and the integration of multi-modality data in medical fields.
% Population-based disease graphs have been successfully constructed in recent research, but their edge selection mostly depends on human design and prior information, often using a single relation type. However, practical applications require automated graph construction methods that consider multiple relations. We propose such a method that also provides transparency and interpretability for clinical understanding. To evaluate the model's performance in medical tasks, we designed two scenarios: Parkinson’s disease (PD) progression and metabolic syndrome prediction.
% PD is a neurodegenerative disorder, resulting in significant deficits in both motor and non-motor functions among patients \citep{armstrong2020diagnosis}. The progression of PD shows considerable heterogeneity, \citep{kalia2015disease, postuma2015mds} and the diversity in the progression of PD presents a complex issue in terms of accurate prediction and disease management, especially in the early stages \citep{frohlich2022leveraging}.
% Metabolic syndrome is a cluster of medical conditions that heighten the risk of developing cardiovascular disease, stroke, and type 2 diabetes. The aforementioned issues include hypertension, hyperglycemia, excessive abdominal obesity, and abnormal levels of cholesterol or triglycerides \citep{nilsson2019metabolic}. The prediction of metabolic syndrome poses challenges due to its intricate nature, encompassing a multifaceted amalgamation of metabolic disorders, including obesity, hyperglycemia, hypertension, and dyslipidemia \citep{micaux2014treatment}.
\begin{figure*}[htb]
\vspace{-3em}
\centering
% Caption and label go in the first argument and the figure contents
% go in the second argument
\floatconts
{fig:Figure1a}
%{\caption{AdaMedGraph: For every feature, a relation graph was bulit and an APPNP model was trained as weak classifier. Utilizing the ADaBoost algorithm, the most optimal weighted weak classifiers were automatically selected. Weighted sum the weak classifiers and get the final predictions.}}
{\setlength{\abovecaptionskip}{0.cm}}{\includegraphics[width=1\linewidth]{images/Figure1b.png}}
{\caption{Overview of AdaMedGraph: (A) we iteratively select $T$ features to train $T$ weak classifiers (i.e., APPNP models), and combine them as an ensemble classifier to provide the final prediction.
(B) Illustration of features propagation within APPNP weak classifier.}}
\vspace{-1em}
\end{figure*}
% The main contributions of our paper are:
% \begin{enumerate}
% \item We propose \ours{} for handling prediction tasks in personalized medicine. \ours{} can automatically construct similarity graph between patients and utilize multi-modal information via GNNs. Inter-individual information and intra-individual features are well-unified in one model, with human efforts on graph building largely relieved.
% \item experiments
% \end{enumerate}
% \textbf{Contribution}
% The main contributions of our paper are:
% \begin{enumerate}
% \item We proposed a multi-modality model leveraging a combination of inter-individual and intra-individual data, allowing for distinct prediction models tailored to each patient, which improved classification accuracy;
% \item Our model consists of automated algorithms for graph construction and takes into account multiple relationships
% \item Our model has the capacity for transparency and interpretability, enabling it to automatically create the significance of patient-relation characteristics and individual-prediction features.
% \end{enumerate}
\section{Related Work}
\label{sec:rwork}
General tabular data processing methods have been widely used to solve medical-related tasks. While these methods based on feature interactions, such as LR and Wide \& Deep, as well as gradient boosting techniques like Gradient Boosting Decision Tree (GBDT) and XGBoost, have shown promise in handling tabular medical data, they tend to focus on feature interactions and often overlook the significance of relationships between patients.
Comparing with general tabular data processing methods, GNNs have the capability to generate embeddings for individual instances by leveraging their inherent information and iteratively gathering messages from neighboring nodes \citep{gilmer2017neural}. Well-established GNN methods encompass GCN \citep{kipf2016semi}, GraphSAGE \citep{hamilton2017inductive}, graph attention networks (GAT) \citep{velickovic2017graph}, and graph isomorphism networks (GIN) \citep{xu2018powerful}. While GNNs excel at capturing intricate node relationships, they are applicable exclusively when dealing with data structured in a graph format. And the quality of the underlying graph structures significantly influences GNN performance \citep{DBLP:journals/corr/abs-2103-03036}. Given that medical tabular data lacks graph structures, the construction of meaningful graphs as inputs for GNNs becomes a critical consideration in this context.
Graph construction for tabular data typically falls into four categories: rule-based, learning-based, search-based, and knowledge-based \citep{li2023graph}.
(1) The rule-based approach operates by either leveraging inherent data dependencies among data instances and features, e.g., \citet{Cvitkovic2020, guo2021tabgnn, you2020handling, zhu2003semi} , or by relying on manually specified heuristics, e.g., \citet{goodge2022lunar, rocheteau2021predicting}. The knowledge-based approach uses domain experts to provide insights into the relationships between data instances, enabling fine-grained graph construction \citep{du2021tabularnet}. These methods require extra heuristics or knowledge.
(2) The learning-based approach for graph construction automatically creates edges between nodes by treating the adjacency matrix as a parameter \citep{hettige2020medgraph, liu2022towards}. However, their internal structures often remain opaque and difficult to understand \citep{xia2021graph}.
%Comparing with them, \ours{} offers another advantage on interpretability, as their internal structures often remain opaque and difficult to understand \citep{xia2021graph}, while \ours{} explicitly selects edge features.
(3) The search-based approach often involves neural architecture search to discover improved graph topologies for representation learning, as demonstrated in \citep{xie2021fives}. However, it prioritizes modeling interactions between features rather than samples or patients. Another search-based approach \citep{du2022learning} models data as a bipartite graph, where features with a specific value are regarded as nodes. Such a graph construction method generally excludes numerical variables from the model.
%\ours{} models interactions between patients, and do not have the above constrain.
Comparing with the current methods, our \ours{} models interactions between patients without the above constraints, provides meaningful automatic graphs constructions, offers an advantage in interpretability, and improves prediction performances.
\section{Methods}
\label{sec:method}
Given $N$ patients, we have the input $X \in \mathbb{R}^{N*M}$ as patient features, and $Y \in \mathbb{R}^{N*K}$ as the one-hot encoding labels.
$M$ is the number of features extracted from multi-model data, and $K$ is the number of label classes.
We propose \ours{} to identify different relationships among patients automatically and then classify the patients accurately.
The whole process is similar to AdaBoost, with GNNs as weak classifiers. In iteration $t$, to specify each weak classifier $g_{\theta_t}$, we need to first select a certain feature $f_t$ to measure the similarity of patients to build an adjacency matrix $A_t$, and then decide the parameters ${\theta_t}$ and $\alpha_t$ by training $g_{\theta_t}(X, A_t)$. The overall objective is to minimize the exponential loss $L(g_T(X), Y)$ where
$$
g_T(X) = \sum_{t=1}^{T} \alpha_t * g_{\theta_t}(X, A_t),
$$
by finding the optimal $\{f_t\}_{t=1,...,T}$, and $\{\theta_t\}_{t=1,...,T}$. $\{\alpha_t\}_{t=1,...,T}$ denote the weights to ensemble GNNs. \figureref{fig:Figure1a} presents the overview of \textit{AdaMedGraph}, and we will dive into the details in the following.
%Given an input graph $\mathcal{G}=\{V,E\}$, where $V$ denote the set of $N$ nodes, and $E$ describe the set of edges. The corresponding adjacency matrix is $A \in R^{N \times N}$. The objective of our methods is to automatically find the best combination of graphs ${G_i}$ where ${i=1,2,...n}$, to make accurate classification predictions for node labels.
%\paragraph{Feature Extraction.}
\paragraph{Initialization}
We first standardize the format of features extracted from multi-modal data by converting them into categorical or numerical values. For example, the images (e.g., MRI) are pre-processed into different brain segments to calculate the volume of grey matter, which serves as an indicator of human cognitive abilities. Additionally, leveraging prior knowledge, genetic data is filtered to retain only the crucial genes associated with the predictions.
At the beginning of the algorithm, we assign the weight of each patient as $w_{i0} = \frac{1}{N}$ where $N$ is the total number of patients. We also specify the type of weak classifiers as Approximate Personalized Propagation of Neural Predictions (APPNP) \citep{klicpera2019predict}, a simple yet effective GNN model tailored for graph-structured data.
\paragraph{Constructing the potential $A_t$}
At iteration $t$, we have the adaptive weights $w_{i,t}$ for data of patient $i$, and the goal is to train a weak classifier $g_{\theta_t}$ and the corresponding weight $\alpha_t$ of $g_{\theta_t}$ to minimize the current weighted loss.
To construct potential adjacency matrices, two things need to be determined:
(1) the feature $f_{j,t}$, $ j \in [1, M]$ that characterizes a specific relationship among patients, and (2) the
threshold $\gamma_t$ that determines the existence of edges in $A_t$.
%threshold $\gamma_t \in [0,1]$ that determines the existence of edges in $A_t$.
Specifically, we explore all the features and consider the 16-quantile, 8-quantile, and 4-quantile of the selected feature values as $\gamma_t$. The edge between patient $i_1$ and $i_2$ is equal to $1$ if the absolute difference between their feature values, $|f_t(i_1) - f_t(i_2) |$, is less than or equal to $\gamma_t$; otherwise, it is set to $0$. In total, we obtain a set of $\{A_t\}$ having $3M$ elements.
% \subsection{Construction of Population Graphs}
% For a single graph, we regard every patient as a node $v$, our proposed methodology requires two design elements: firstly, the choice of the feature vector $\operatorname{x}(v)$ that characterizes each sample, and secondly, the choice of interactions between samples, accomplished through the definition of the graph edges $E$. We employ all available input information $F={f_i}$ as node features and systematically explore each input feature $f_i$ to decide the construct population graphs through an iterative process. Specifically, we define a similarity function $Sim$, and threshold $\theta_i$, to decide the connection of patient v and w using feature $f_i$:
% \begin{equation}\label{eq:eq1}
% Sim (f_i^v,f_i^w )=
% \left\{\begin{array}{l}
% 1, \text{ if } |f_i^v-f_i^w |< \theta_i,\\
% 0, \text{ otherwise} \\
% \end{array}
% \right.
% \end{equation}
% In this case, we can build a set of population graph $\{G_i\}$ where each graph represents a kind of relationship between patients.
\paragraph{Training the weak classifier $g_{\theta_t}$}
Next, we aim to find an optimal choice of $A_t$, $\theta_t$ and $\alpha_t$ satisfying
$$
\min_{A_t, \theta_t, \alpha_t} L( g_{\theta_t}(X, A_t), Y) = \sum_{i=1}^{N} w_{i,t} \cdot \exp(- \alpha_t \cdot y_i \cdot g_{\theta_t}(x_i, A_t)).
$$
Following the SAMME \citep{zhu2006multi}, the optimal $\theta_t$ is equal to minimize the error at iteration $t$
$$ \small
\text{err}_t = \sum_{i=1}^{N} w_{i,t} \cdot \mathds{1}(y_i \neq g_{\theta_t}(x_i, A_t)),
$$
and the corresponding weight $\alpha_t$ is calculated by
$$ \small
\alpha_t = \frac{1}{2}\log \frac{1-\text{err}_t}{\text{err}_t} + \log(K-1).
$$
For every $A_t$, we can train an APPNP model to obtain $\theta_t$ and $\alpha_t$. Thus, we explore each possible value in the previous constructed $\{A_t\}$ to find a minimal $ L( g_{\theta_t}(X, A_t), Y)$.
After determining the optimal solution, the weights of the patients are updated by
$$
w_{i,t+1} = w_{i,t} \cdot \exp(\alpha_t \cdot \mathds{1}(y_i \neq g_{\theta_t}(x_i, A_t) ),
$$
and we proceed to the next iteration.
\paragraph{Termination}
The whole process is ended at the iteration $T$ when the error rate of $g_T(X)$ or $g_{\theta_T}(X, A_T)$ gets equal or larger than $\frac{K-1}{K}$.
% \subsection{AdaBoost with Graph Neural Network as weak classifier}
% The AdaMedGraph framework leverages the power of Graph Neural Network (GNN) \citep{scarselli2008graph} by aggregating them using the AdaBoost \citep{hastie2009multi} algorithm to enhance predictive performance. Specifically, the AdaMedGraph classifier framework can be write as:
% \begin{equation}\label{eq:eq2}
% F(x(v))= \sum_{i=1}^{M}\alpha_i*g_i (x(v),G_i)
% \end{equation}
% where $F(x(v))$is the ensemble classifier obtained after m rounds of training, and $x(v)$ denotes node features. For each round, the framework automatically adds a trained GNN classifier $g_i (x(v),G_i)$ to improve the prediction performance. The weight of the classifier $\alpha_i$ denotes the importance of classifier (also the kind of relation between patients), and it could be obtained according to the error of the classifier during training. See \figureref{fig:Figure1a} for the overview of our proposed model.
\begin{table*}[]
\vspace{-4em}
\caption{PPMI and PDBP 24-month prediction AUROC score. M* is MDS-UPDRS.}
\label{tab:PPMIPDBP}
\scalebox{0.8}{
\begin{tabular}{lllllllllllll}
\hline
& \multicolumn{2}{c}{\textbf{MLP}} & \multicolumn{2}{c}{\textbf{LR}} & \multicolumn{2}{c}{\textbf{SVM}} & \multicolumn{2}{c}{\textbf{RF}} & \multicolumn{2}{c}{\textbf{XGB}} & \multicolumn{2}{c}{\textbf{AdaMedGraph}} \\ \hline
Label & PPMI & PDBP & PPMI & PDBP & PPMI & PDBP & PPMI & PDBP & PPMI & PDBP & PPMI & PDBP \\ \hline
HY & 0.593 & 0.450 & 0.652 & 0.504 & 0.705 & 0.614 & 0.664 & 0.692 & 0.717 & 0.643 & \textbf{0.775} & \textbf{0.701} \\ \hline
\begin{tabular}[c]{@{}l@{}}M* I\end{tabular} & 0.445 & 0.473 & 0.601 & 0.541 & 0.573 & 0.554 & 0.605 & 0.581 & 0.625 & 0.584 & \textbf{0.656} & \textbf{0.586} \\ \hline
\begin{tabular}[c]{@{}l@{}}M* II\end{tabular} & 0.441 & 0.499 & 0.578 & 0.536 & 0.570 & 0.610 & 0.602 & 0.598 & 0.620 & 0.602 & \textbf{0.651} & \textbf{0.621} \\ \hline
\begin{tabular}[c]{@{}l@{}}M* III\end{tabular} & 0.695 & 0.645 & 0.691 & 0.648 & 0.695 & 0.645 & 0.670 & 0.643 & 0.710 & 0.610 & \textbf{0.717} & \textbf{0.660} \\ \hline
\begin{tabular}[c]{@{}l@{}}M* Total\end{tabular} & 0.615 & 0.577 & 0.635 & 0.587 & 0.634 & 0.598 & 0.595 & 0.579 & 0.613 & 0.590 & \textbf{0.646} & \textbf{0.601} \\ \hline
\begin{tabular}[c]{@{}l@{}}M* Axial\end{tabular} & 0.618 & 0.552 & 0.637 & 0.653 & 0.661 & 0.652 & 0.645 & 0.657 & 0.627 & 0.618 & \textbf{0.661} & \textbf{0.660} \\ \hline
\begin{tabular}[c]{@{}l@{}}M* Rigid\end{tabular} & 0.644 & 0.659 & 0.665 & 0.656 & 0.654 & 0.678 & 0.663 & 0.658 & 0.591 & 0.626 & \textbf{0.694} & \textbf{0.683} \\ \hline
\begin{tabular}[c]{@{}l@{}}M* Tremor\end{tabular} & 0.669 & 0.651 & 0.689 & 0.635 & 0.684 & 0.619 & 0.713 & 0.625 & 0.672 & 0.613 & \textbf{0.715} & \textbf{0.685} \\ \hline
MoCA & 0.519 & 0.497 & 0.600 & \textbf{0.694} & 0.574 & 0.559 & 0.573 & 0.623 & 0.614 & 0.684 & \textbf{0.660} & 0.506 \\ \hline
ESS & 0.573 & 0.546 & 0.573 & 0.630 & 0.601 & 0.628 & 0.625 & 0.625 & 0.611 & 0.628 & \textbf{0.645} & \textbf{0.631} \\ \hline
\end{tabular}
}
\end{table*}
\section{Data and Experiments}
\label{sec:experiments}
% We implemented PD progression and metabolic syndrome incidence prediction experiments.
In this section, we first introduce prediction tasks and settings in two real-world medical scenarios and then present the results of these tasks.
\subsection{PD progression}
The longitudinal data from the Parkinson's Progression Markers Initiative (PPMI) \citep{marek2018parkinson} and the Parkinson's Disease Biomarkers Program (PDBP) \citep{gwinn2017parkinson}are used in our work.
\paragraph{Prediction tasks}
We select 10 assessment scores as labels for disease progression: Movement Disorder Society-sponsored Revision of the Unified Parkinson's Disease Rating Scale (MDS-UPDRS) I, II, III, and total (sum of I, II, and III), and the sub-scores for tremors, axial symptoms, and rigidity in MDS-UPDRS III; Hoehn and Yahr Scale (HY); Montreal Cognitive Assessment (MoCA) total; and Epworth Sleepiness Scale (ESS) total. To get the classification labels, we compute the mean and standard deviation of changes in each score (except HY since HY has a standard classification scale) from baseline to 24-months in the healthy control (HC) group and compare them with the changes in the PD group. By constructing a reference interval based on the HC changes, we then categorize the PD changes as ``No Change" if within the interval, or as ``Better" or ``Worse" if less or more than the interval, respectively.
\paragraph{Feature extraction}
We pre-possess genetic data and magnetic resonance imaging (MRI) before training. For MRI, the participant's T1 scan has been processed into 138 segments, and the volumes of the segments are used as input. For genetic data, the monogenic mutation statuses of GBA, LRRK2, SNCA, and APOE\_E4 are encoded as binary variables.
\paragraph{Dataset statistics and division}
Two sets of experiments have been designed for the PD progression task. In PD Experiment I, we utilize clinical assessments and genetic data as input features, comprising a total of 78 variables, given the absence of MRI data in the PDBP dataset. The PPMI dataset is split into a training set (80\%, n = 358) and an internal testing set (20\%, n = 90). Besides, we randomly select 20\% of the training data as our validation set during training to select the final best models. The PDBP dataset serves as an external validation dataset (n = 228).
In Experiment II, we expand the range of input features to include clinical assessments, MRI data, and genetic information, resulting in a total of 234 variables. This study exclusively utilizes the PPMI dataset, which has been divided into an 80\% training set (n = 252) and a 20\% testing set (n = 65). Similarly, 20\% of training data has been randomly separated as validation set. In both experiments, the train-test split is based on participants' enrollment times, with the earlier visit participants assigned to the training set while the later visit participants are allocated to the testing set, while the validation split is done randomly.
\paragraph{Baselines}
To evaluate the performance of our model, we include several ML classification methods as baseline models. The aforementioned models have frequently been employed in medical classification tasks, encompassing a 2-layer Multilayer Perceptron (MLP), a logistic regression classifier (LR), a random forest classifier (RF), a support vector machine classifier (SVM), and a XGBoost model.
Furthermore, in order to comprehend the importance of automatically selecting edge features, we construct a graph based on prior knowledge. Since it is well-known that age is a critical factor for PD progression, we construct a graph based on age (with a threshold of 5) and develop an APPNP (referred to as APPNP-Age) as a comparative model in experiment II.
\subsection{Metabolic syndrome}
In the metabolic syndrome prediction task, we include 109,027 individuals in the UK Biobank \citep{sudlow2015uk} (80\% for training and 20\% for testing). All subjects have no record of metabolic syndrome on their baseline screen, and their 168 proton nuclear magnetic resonance (NMR) metabolic biomarkers and 17 traditional clinical risk variables are used as input \citep{lian2023metabolic} to predict whether participants will have the diseases after a period of time. The input data has been processed into 9 principal components using principal component analysis, and XGBoost and TabNet \citep{arik2021tabnet} are implemented as baseline models.
% to compare with \ours{}.
\subsection{Results}
\label{sec:result}
% We utilize the weighted area under the receiver operating characteristic curve (AUROC) for multiclass task \citep{bishop2006pattern} as the evaluation metric for PD task and AUROC for Metabolic syndrome.
We utilize the weighted area under the receiver operating characteristic curve (AUROC) as the evaluation metric for all tasks.
\paragraph{PD progression}
In the context of predicting 24-month progression experiment I, our AdaMedGraph model has received greater accuracy in comparison to all baseline models across all labels, with the exception of MoCA in the PDBP dataset. This discrepancy can be attributed to differences in the distribution of selected edge features for MoCA between the two datasets. These superior results from our model underscore the significance of incorporating both intra- and inter-patient data for individual disease progression prediction. Refer to the \tableref{tab:PPMIPDBP} for a comprehensive understanding of the performances.
% In terms of 24-month HY prediction, AdaMedGraph outperformed XGBoost, RF, SVM, LR, and MLP with AUC values of 0.775 (PPMI) and 0.701 (PDBP), compared to the AUC values of 0.717 and 0.643 for XGBoost, 0.664 and 0.692 for RF, 0.705 and 0.614 for SVM, 0.652 and 0.504 for LR, and 0.593 and 0.450 for MLP.
In experiment II, AdaMedGraph also exhibits superior performance when compared to all the baseline models as well as APPNP-age, which suggests that taking into account automatically selected relationships among patients can significantly enhance the performance of prediction. Please refer to \tableref{tab:PPMI} for details.
\begin{table}[h]
\caption{PPMI 24-month prediction AUROC score. A-A is APPNP-Age. M* is MDS-UPDRS.}
\label{tab:PPMI}
\scalebox{0.75}{
\begin{tabular}{llllllll}
\hline
Label & MLP & LR & SVM & RF & XGB & A-A & Ours \\ \hline
HY & 0.556 & 0.630 & 0.594 & 0.633 & 0.643 & 0.630 & \textbf{0.682} \\ \hline
\begin{tabular}[c]{@{}l@{}}M* I\end{tabular} & 0.543 & 0.638 & 0.531 & 0.593 & 0.628 & 0.699 & \textbf{0.730} \\ \hline
\begin{tabular}[c]{@{}l@{}}M* II\end{tabular} & 0.507 & 0.588 & 0.573 & 0.457 & 0.575 & 0.621 & \textbf{0.652} \\ \hline
\begin{tabular}[c]{@{}l@{}}M* III\end{tabular} & 0.409 & 0.583 & 0.587 & 0.553 & 0.566 & 0.604 & \textbf{0.666} \\ \hline
\begin{tabular}[c]{@{}l@{}}M* Total\end{tabular} & 0.648 & 0.617 & 0.593 & 0.590 & 0.598 & 0.563 & \textbf{0.684} \\ \hline
\begin{tabular}[c]{@{}l@{}}M* Axial\end{tabular} & 0.517 & 0.610 & 0.610 & 0.549 & 0.538 & 0.591 & \textbf{0.689} \\ \hline
\begin{tabular}[c]{@{}l@{}}M* Rigid\end{tabular} & 0.419 & 0.572 & 0.636 & 0.548 & 0.628 & 0.682 & \textbf{0.693} \\ \hline
\begin{tabular}[c]{@{}l@{}}M* Tremor\end{tabular} & 0.520 & 0.628 & 0.648 & 0.657 & \textbf{0.700} & 0.585 & \textbf{0.700} \\ \hline
MoCA & 0.680 & 0.713 & 0.672 & 0.714 & 0.723 & 0.668 & \textbf{0.746} \\ \hline
ESS & 0.537 & 0.623 & 0.639 & 0.645 & 0.596 & 0.649 & \textbf{0.676} \\ \hline
\end{tabular}}
\end{table}
\vspace{-10pt}
\paragraph{Metabolic syndrome}
In the context of predicting the occurrence of metabolic syndrome, the AdaMedGraph model demonstrates superior performance with an area under the AUROC of 0.675 on the testing dataset. This surpasses both the XGBoost model, which achieved an AUROC of 0.641, and the TabNet model, which achieved an AUROC of 0.672.
\section{Discussion}
\label{sec:discussion}
In this study, we introduce an innovative algorithm, namely \ours{}, designed to autonomously identify important features for the construction of multiple patient similarity graphs, which serve as the basis for training GNNs in an AdaBoost framework, thereby enhancing the accuracy of classification tasks. We have conducted two sets of clinical experiments, and the results affirm our initial hypothesis that automatically constructing multi-relationship graphs among patients using inter- and intra-individual data can benefit personalized medicine.
One thing that needs to be noticed is the computational cost of our \ours{}. Although our model goes through all potential features to construct a total of $3M$ graphs during training, it's important to note that we use the APPNP model as the weak classifier, known for its relatively low computational cost, with only linear computational complexity. In this case, the total computational cost of our model is acceptable. In our experiments, we have observed reasonable training time costs. For instance, in our PD prediction task with 234 input features and 317 patient samples, the training process has taken less than 5 minutes. Similarly, for the metabolic syndrome prediction task, which includes 109,027 patients and 9 features, it takes less than 30 minutes for training. These experiments have been conducted on a single V100 GPU. Furthermore, to optimize for high-dimensional settings, a simple feature selection process could be integrated before searching for the best features.
% \begin{table}[htb]
% \centering
% \setlength{\tabcolsep}{14mm}
% {
% \begin{tabular}{@{}ll@{}}
% \toprule
% Model & AUC \\ \midrule
% XGBoost & 0.641 \\
% TabNet & 0.671 \\
% AdaMedGraph & 0.675 \\ \bottomrule
% \end{tabular}
% }
% \caption{Metabolic syndrome prediction performance of XGBoost, TabNet and AdaMedGraph}
% \label{tab:mate}
% \end{table}
% \acks{}
\bibliography{jmlr-sample}
\newpage
\appendix
\section{Ethics Statement}\label{apd:first}
The population cohort in the metabolic syndrome experiments is from the UK Biobank [Application Number 78730] which has received ethical approval from the North West Multicentre Research Ethics Committee (REC reference: 11/NW/03820). All participants have given written informed consent before enrolment.
PD study data is retrieved from AMPPD platform https://amp-pd.org/ using the terms of the AMP-PD Data Use Agreement and PPMI website https://www.ppmi-info.org/. The data has used in this study has been collected before the commencement of our study and has been obtained in an anonymized format. In the original PPMI and PDBP studies, participants have previously consented in writing to the sharing of their data. These investigations have been conducted in accordance with protocols authorized by the Indiana University Institutional Review Board (IRB) for PPMI and the respective PDBP centers.
% This is the first appendix.
\section{Hyper-parameters}\label{apd:first}
We list the hyper-parameters of baseline models and \ours{} in this part. Grid search method has used for all models.
\subsection{Task 1}
\paragraph{MLP} We train the MLP models with Adam optimizer and search the hyper-parameters of hidden dimension in \{128, 256, 512, 1024\}, L2 regularization in \{$10^{-5}$, $10^{-4}$, $10^{-3}$\}, learning rate \{$10^{-5}$, $5*10^{-4}$, $10^{-4}$, $5*10^{-3}$\}, dropout in \{0.1, 0.3, 0.5\} and total epoch in \{100, 200, 300\}.
\paragraph{LR} We search the LR models hyper-parameters of regularization method in \{L1, L2, elastic net\} and strength of the regularization in \{0.01, 0.1, 1, 10\}.
\paragraph{SVM} We search the the LR models hyper-parameters of kernels in \{linear, polynomial, radial basis function\} and strength of the regularization in in \{0.01, 0.1, 1, 10\}.
\paragraph{RF} We search the LR models hyper-parameters of number of estimators in \{50, 100, 150, 200 \}, max depth in \{10, 20, 30, 40, 50\}, minimum number of samples required to be at a leaf node in \{1, 2, 4\}, and the minimum number of samples required to split an internal node in \{1, 2, 3\}.
\paragraph{XGB} We list the XGB models hyper-parameters searching in \tableref{tab:xgbtable}.
\begin{table}[h]
\caption{XGBoost hyper-parameters search list}
\label{tab:xgbtable}
\begin{tabular}{|l|l|lll}
\cline{1-2}
Hyper-parameter & Tunning Range & & & \\ \cline{1-2}
Number of estimators & \{50, 75, 100, 150, 200\} & & & \\ \cline{1-2}
Max depth & \{1, 2, 3, 6, 9\} & & & \\ \cline{1-2}
Learning rate & \{0.001, 0.01, 0.1, 0.3, 0.5\} & & & \\ \cline{1-2}
Min child weight & \{1, 2, 3, 4, 5\} & & & \\ \cline{1-2}
Gamma & \{ 0.5, 1, 2, 3, 6, 9\} & & & \\ \cline{1-2}
Colsample by tree & \{0, 0.25, 0.5, 0.75, 1\} & & & \\ \cline{1-2}
Colsample by level & \{0, 0.25, 0.5, 0.75, 1\} & & & \\ \cline{1-2}
lambda & \{0.05, 0.1, 0.25, 0.5, 1\} & & & \\ \cline{1-2}
alpha & \{0.5, 1, 10\} & & & \\ \cline{1-2}
\end{tabular}
\end{table}
\paragraph{APPNP-Age} We train our APPNP-Age with Adam optimizer, Cosine Annealing Scheduler for weight decay, and 100 total epochs. We search the hyper-parameters of hidden dimension of linear layer (H) in \{128, 256, 512\}, number of iterations (k) in \{3, 5, 10\}, teleport probability in \{0.1, 0.3, 0.5, 1\}, dropout in \{0.1, 0.3, 0.5\}, learning rate in \{$10^{-5}$, $5*10^{-4}$, $10^{-4}$, $5*10^{-3}$ \}, and threshold for age in \{4, 4.5, 5, 5.5, 6\}.
\paragraph{AdaMedGraph} The training details for AdaMedGraph contain two parts: training a single APPNP and training AdaBosst. The details have been summarized in \tableref{tab:adamed}.
\begin{table}[h]
\caption{AdaMedGraph hyper-parameter list. Hyper-parameters in braces are searched.}
\label{tab:adamed}
\scalebox{0.95}{
\begin{tabular}{|ll|}
\hline
\multicolumn{2}{|c|}{Single APPNP} \\ \hline
\multicolumn{1}{|l|}{Optimizer} & Adam \\ \hline
\multicolumn{1}{|l|}{Weight decay} & \{0.00001, 0.0001, 0.001\} \\ \hline
\multicolumn{1}{|l|}{Learning rate decay} & Cosine Annealing Scheduler \\ \hline
% \multicolumn{1}{|l|}{Learning rate} & \{$10^{-5}$, $5*10^{-5}$, $10^{-4}$, $5*10^{-4}$, $10^{-3}$, $5*10^{-3}$\} \\ \hline
\multicolumn{1}{|l|}{Learning rate} & \begin{tabular}[c]{@{}l@{}}\{$10^{-5}$, $5*10^{-5}$, $10^{-4}$, \\ $5*10^{-4}$, $10^{-3}$, $5*10^{-3}$\}\end{tabular} \\ \hline
\multicolumn{1}{|l|}{Total epoch} & \{100, 150, 200\} \\ \hline
\multicolumn{1}{|l|}{Early stop} & \{5, 10, 20\} \\ \hline
\multicolumn{1}{|l|}{H} & \{128, 256, 512\} \\ \hline
\multicolumn{1}{|l|}{k} & \{3, 5, 10\} \\ \hline
\multicolumn{1}{|l|}{Teleport probability} & \{0.1, 0.3, 0.5, 1\} \\ \hline
\multicolumn{1}{|l|}{Dropout} & \{0.1, 0,3, 0,5\} \\ \hline
\multicolumn{2}{|c|}{AdaBoost} \\ \hline
\multicolumn{1}{|l|}{Number of estimators} & \{5, 10, 15, 20\} \\ \hline
\multicolumn{1}{|l|}{Learning rate} & \{0.1, 0.33, 0.5, 0.66, 1\} \\ \hline
\end{tabular}}
\end{table}
\subsection{Task 2: metabolic syndrome prediction}
\paragraph{XGB} We list the XGB models hyper-parameters searching in \tableref{tab:xgbtable2}.
\vspace{-2em}
\begin{table}[h]
\caption{XGBoost hyper-parameters search list for metabolic syndrome prediction}
\label{tab:xgbtable2}
\begin{tabular}{|l|l|lll}
\cline{1-2}
Hyper-parameter & Tunning Range & & & \\ \cline{1-2}
Number of estimators & \{50, 100, 200, 300\} & & & \\ \cline{1-2}
Max depth & \{3, 6, 9, 12\} & & & \\ \cline{1-2}
Learning rate & \{0.001, 0.01, 0.1, 0.3, 0.5\} & & & \\ \cline{1-2}
Min child weight & \{1, 2, 3, 4, 5\} & & & \\ \cline{1-2}
Gamma & \{ 0.5, 1, 3, 6, 9, 12\} & & & \\ \cline{1-2}
Colsample by tree & \{0, 0.5, 1\} & & & \\ \cline{1-2}
Colsample by level & \{0, 0.5, 1\} & & & \\ \cline{1-2}
lambda & \{0.05, 0.1, 0.25, 0.5, 1\} & & & \\ \cline{1-2}
alpha & \{0.001, 0.01, 0.5, 1, 10\} & & & \\ \cline{1-2}
\end{tabular}
\end{table}
\paragraph{TabNet} We Search the TabNet models hyper-parameters of mask type in \{”entmax”, ”sparsemax”\}, the width of the decision prediction layer and the attention embedding for each mask in \{32, 48, 56, 64\}, number of steps in \{1, 2, 3\}, gamma in \{1.0, 1.2, 1.4\}, the number of shared Gated Linear Units at each step in \{1, 2, 3\}, learning rate in \{$10^{-5}$, $5*10^{-5}$, $10^{-4}$, $5*10^{-4}$, $10^{-3}$, $5*10^{-3}$\} and lambda space in \{$10^{-5}$, $10^{-4}$\}.
\paragraph{AdaMedGraph} The training details for AdaMedGraph on task 2 have been summarized in \tableref{tab:adamed2}.
\begin{table}[!h]
\vspace{-2em}
\caption{AdaMedGraph hyper-parameter for metabolic syndrome prediction. Hyper-parameters in braces are searched.}
\label{tab:adamed2}
\scalebox{0.95}{
\begin{tabular}{|ll|}
\hline
\multicolumn{2}{|c|}{Single APPNP} \\ \hline
\multicolumn{1}{|l|}{Optimizer} & Adam \\ \hline
\multicolumn{1}{|l|}{Weight decay} & \{0.00001, 0.0001, 0.001\} \\ \hline
\multicolumn{1}{|l|}{Learning rate decay} & Cosine Annealing Scheduler \\ \hline
% \multicolumn{1}{|l|}{Learning rate} & \{$10^{-5}$, $5*10^{-5}$, $10^{-4}$, $5*10^{-4}$, $10^{-3}$, $5*10^{-3}$\} \\ \hline
\multicolumn{1}{|l|}{Learning rate} & \begin{tabular}[c]{@{}l@{}}\{$10^{-5}$, $5*10^{-5}$, $10^{-4}$, \\ $5*10^{-4}$, $10^{-3}$, $5*10^{-3}$\}\end{tabular} \\ \hline
\multicolumn{1}{|l|}{Total epoch} & \{100, 200, 500, 1000\} \\ \hline
\multicolumn{1}{|l|}{Early stop} & \{10, 20, 30\} \\ \hline
\multicolumn{1}{|l|}{H} & \{128, 256, 512\} \\ \hline
\multicolumn{1}{|l|}{k} & \{3, 5, 10\} \\ \hline
\multicolumn{1}{|l|}{Teleport probability} & \{0.1, 0.3, 0.5, 1\} \\ \hline
\multicolumn{1}{|l|}{Dropout} & \{0.1, 0,3, 0,5\} \\ \hline
\multicolumn{2}{|c|}{AdaBoost} \\ \hline
\multicolumn{1}{|l|}{Number of estimators} & \{5, 10, 15, 20\} \\ \hline
\multicolumn{1}{|l|}{Learning rate} & \{0.1, 0.33, 0.5, 0.66, 1\} \\ \hline
\end{tabular}}
\end{table}
% This is the second appendix.
\end{document}
\begin{thebibliography}{34}
\providecommand{\natexlab}[1]{#1}
\providecommand{\url}[1]{\texttt{#1}}
\expandafter\ifx\csname urlstyle\endcsname\relax
\providecommand{\doi}[1]{doi: #1}\else
\providecommand{\doi}{doi: \begingroup \urlstyle{rm}\Url}\fi
\bibitem[Arik and Pfister(2021)]{arik2021tabnet}
Sercan~{\"O} Arik and Tomas Pfister.
\newblock Tabnet: Attentive interpretable tabular learning.
\newblock In \emph{Proceedings of the AAAI conference on artificial
intelligence}, volume~35, pages 6679--6687, 2021.
\bibitem[Cvitkovic(2020)]{Cvitkovic2020}
Milan Cvitkovic.
\newblock Supervised learning on relational databases with graph neural
networks.
\newblock \emph{International Conference on Learning Representations}, 2020.
\bibitem[Du et~al.(2022)Du, Zhang, Zhou, Wang, Zhao, Jin, Gan, Zhang, and
Wipf]{du2022learning}
Kounianhua Du, Weinan Zhang, Ruiwen Zhou, Yangkun Wang, Xilong Zhao, Jiarui
Jin, Quan Gan, Zheng Zhang, and David~P Wipf.
\newblock Learning enhanced representation for tabular data via neighborhood
propagation.
\newblock \emph{Advances in Neural Information Processing Systems},
35:\penalty0 16373--16384, 2022.
\bibitem[Du et~al.(2021)Du, Gao, Chen, Jia, Wang, Zhang, Han, and
Zhang]{du2021tabularnet}
Lun Du, Fei Gao, Xu~Chen, Ran Jia, Junshan Wang, Jiang Zhang, Shi Han, and
Dongmei Zhang.
\newblock Tabularnet: A neural network architecture for understanding semantic
structures of tabular data.
\newblock In \emph{Proceedings of the 27th ACM SIGKDD Conference on Knowledge
Discovery \& Data Mining}, pages 322--331, 2021.
\bibitem[Gilmer et~al.(2017)Gilmer, Schoenholz, Riley, Vinyals, and
Dahl]{gilmer2017neural}
Justin Gilmer, Samuel~S Schoenholz, Patrick~F Riley, Oriol Vinyals, and
George~E Dahl.
\newblock Neural message passing for quantum chemistry.
\newblock In \emph{International conference on machine learning}, pages
1263--1272. PMLR, 2017.
\bibitem[Goodge et~al.(2022)Goodge, Hooi, Ng, and Ng]{goodge2022lunar}
Adam Goodge, Bryan Hooi, See-Kiong Ng, and Wee~Siong Ng.
\newblock Lunar: Unifying local outlier detection methods via graph neural
networks.
\newblock In \emph{Proceedings of the AAAI Conference on Artificial
Intelligence}, volume~36, pages 6737--6745, 2022.
\bibitem[Guo et~al.(2021)Guo, Quan, Zhao, Yao, Li, and Tu]{guo2021tabgnn}
Xiawei Guo, Yuhan Quan, Huan Zhao, Quanming Yao, Yong Li, and Weiwei Tu.
\newblock Tabgnn: Multiplex graph neural network for tabular data prediction.
\newblock \emph{The 3rd Workshop on Deep Learning Practice for High-Dimensional
Sparse Data with KDD (DLP-KDD)}, 2021.
\bibitem[Gwinn et~al.(2017)Gwinn, David, Swanson-Fischer, Albin,
Hillaire-Clarke, Sieber, Lungu, Bowman, Alcalay, Babcock,
et~al.]{gwinn2017parkinson}
Katrina Gwinn, Karen~K David, Christine Swanson-Fischer, Roger Albin, Coryse~St
Hillaire-Clarke, Beth-Anne Sieber, Codrin Lungu, F~DuBois Bowman, Roy~N
Alcalay, Debra Babcock, et~al.
\newblock Parkinson's disease biomarkers: perspective from the {NINDS}
{P}arkinson's disease biomarkers program.
\newblock \emph{Biomarkers in medicine}, 11\penalty0 (6):\penalty0 451--473,
2017.
\bibitem[Hamilton et~al.(2017)Hamilton, Ying, and
Leskovec]{hamilton2017inductive}
Will Hamilton, Zhitao Ying, and Jure Leskovec.
\newblock Inductive representation learning on large graphs.
\newblock \emph{Advances in neural information processing systems}, 30, 2017.
\bibitem[Hettige et~al.(2020)Hettige, Wang, Li, Le, and
Buntine]{hettige2020medgraph}
Bhagya Hettige, Weiqing Wang, Yuan-Fang Li, Suong Le, and Wray Buntine.
\newblock Medgraph: Structural and temporal representation learning of
electronic medical records.
\newblock In \emph{ECAI 2020}, pages 1810--1817. IOS Press, 2020.
\bibitem[Hulsen et~al.(2019)Hulsen, Jamuar, Moody, Karnes, Varga, Hedensted,
Spreafico, Hafler, and McKinney]{hulsen2019big}
Tim Hulsen, Saumya~S Jamuar, Alan~R Moody, Jason~H Karnes, Orsolya Varga, Stine
Hedensted, Roberto Spreafico, David~A Hafler, and Eoin~F McKinney.
\newblock From big data to precision medicine.
\newblock \emph{Frontiers in medicine}, 6:\penalty0 34, 2019.
\bibitem[Kipf and Welling(2016)]{kipf2016semi}
Thomas~N Kipf and Max Welling.
\newblock Semi-supervised classification with graph convolutional networks.
\newblock In \emph{International Conference on Learning Representations}, 2016.
\bibitem[Klicpera et~al.(2019)Klicpera, Bojchevski, and
G{\"u}nnemann]{klicpera2019predict}
Johannes Klicpera, Aleksandar Bojchevski, and Stephan G{\"u}nnemann.
\newblock Predict then propagate: Graph neural networks meet personalized
pagerank.
\newblock 2019.
\bibitem[Li et~al.(2023)Li, Tsai, and Liao]{li2023graph}
Cheng-Te Li, Yu-Che Tsai, and Jay~Chiehen Liao.
\newblock Graph neural networks for tabular data learning.
\newblock In \emph{2023 IEEE 39th International Conference on Data Engineering
(ICDE)}, pages 3589--3592. IEEE, 2023.
\bibitem[Lian and Vardhanabhuti(2023)]{lian2023metabolic}
Jie Lian and Varut Vardhanabhuti.
\newblock Metabolic biomarkers using nuclear magnetic resonance metabolomics
assay for the prediction of aging-related disease risk and mortality: a
prospective, longitudinal, observational, cohort study based on the uk
biobank.
\newblock \emph{GeroScience}, pages 1--12, 2023.
\bibitem[Liu et~al.(2022)Liu, Zheng, Zhang, Chen, Peng, and
Pan]{liu2022towards}
Yixin Liu, Yu~Zheng, Daokun Zhang, Hongxu Chen, Hao Peng, and Shirui Pan.
\newblock Towards unsupervised deep graph structure learning.
\newblock In \emph{Proceedings of the ACM Web Conference 2022}, pages
1392--1403, 2022.
\bibitem[Marek et~al.(2018)Marek, Chowdhury, Siderowf, Lasch, Coffey,
Caspell-Garcia, Simuni, Jennings, Tanner, Trojanowski,
et~al.]{marek2018parkinson}
Kenneth Marek, Sohini Chowdhury, Andrew Siderowf, Shirley Lasch, Christopher~S
Coffey, Chelsea Caspell-Garcia, Tanya Simuni, Danna Jennings, Caroline~M
Tanner, John~Q Trojanowski, et~al.
\newblock The {P}arkinson's progression markers initiative
({PPMI})--establishing a {PD} biomarker cohort.
\newblock \emph{Annals of clinical and translational neurology}, 5\penalty0
(12):\penalty0 1460--1477, 2018.
\bibitem[Mhasawade et~al.(2021)Mhasawade, Zhao, and
Chunara]{mhasawade2021machine}
Vishwali Mhasawade, Yuan Zhao, and Rumi Chunara.
\newblock Machine learning and algorithmic fairness in public and population
health.
\newblock \emph{Nature Machine Intelligence}, 3\penalty0 (8):\penalty0
659--666, 2021.
\bibitem[Parisot et~al.(2017)Parisot, Ktena, Ferrante, Lee, Moreno, Glocker,
and Rueckert]{parisot2017spectral}
Sarah Parisot, Sofia~Ira Ktena, Enzo Ferrante, Matthew Lee, Ricardo~Guerrerro
Moreno, Ben Glocker, and Daniel Rueckert.
\newblock Spectral graph convolutions for population-based disease prediction.
\newblock In \emph{Medical Image Computing and Computer Assisted Intervention-
MICCAI 2017: 20th International Conference, Quebec City, QC, Canada,
September 11-13, 2017, Proceedings, Part III 20}, pages 177--185. Springer,
2017.
\bibitem[Rocheteau et~al.(2021)Rocheteau, Tong, Veli{\v{c}}kovi{\'c}, Lane, and
Li{\`o}]{rocheteau2021predicting}
Emma Rocheteau, Catherine Tong, Petar Veli{\v{c}}kovi{\'c}, Nicholas Lane, and
Pietro Li{\`o}.
\newblock Predicting patient outcomes with graph representation learning.
\newblock \emph{International Workshop on Deep Learning on Graphs: Method and
Applications (DLG-AAAI)}, 2021.
\bibitem[Shilo et~al.(2020)Shilo, Rossman, and Segal]{shilo2020axes}
Smadar Shilo, Hagai Rossman, and Eran Segal.
\newblock Axes of a revolution: challenges and promises of big data in
healthcare.
\newblock \emph{Nature medicine}, 26\penalty0 (1):\penalty0 29--38, 2020.
\bibitem[Sudlow et~al.(2015)Sudlow, Gallacher, Allen, Beral, Burton, Danesh,
Downey, Elliott, Green, Landray, et~al.]{sudlow2015uk}
Cathie Sudlow, John Gallacher, Naomi Allen, Valerie Beral, Paul Burton, John
Danesh, Paul Downey, Paul Elliott, Jane Green, Martin Landray, et~al.
\newblock Uk biobank: an open access resource for identifying the causes of a
wide range of complex diseases of middle and old age.
\newblock \emph{PLoS medicine}, 12\penalty0 (3):\penalty0 e1001779, 2015.
\bibitem[Tu et~al.(2023)Tu, Azizi, Driess, Schaekermann, Amin, Chang, Carroll,
Lau, Tanno, Ktena, et~al.]{tu2023towards}
Tao Tu, Shekoofeh Azizi, Danny Driess, Mike Schaekermann, Mohamed Amin,
Pi-Chuan Chang, Andrew Carroll, Chuck Lau, Ryutaro Tanno, Ira Ktena, et~al.
\newblock Towards generalist biomedical {AI}.
\newblock \emph{arXiv preprint arXiv:2307.14334}, 2023.
\bibitem[Velickovic et~al.(2017)Velickovic, Cucurull, Casanova, Romero, Lio,
Bengio, et~al.]{velickovic2017graph}
Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro
Lio, Yoshua Bengio, et~al.
\newblock Graph attention networks.
\newblock \emph{stat}, 1050\penalty0 (20):\penalty0 10--48550, 2017.
\bibitem[Xia et~al.(2021)Xia, Sun, Yu, Aziz, Wan, Pan, and Liu]{xia2021graph}
Feng Xia, Ke~Sun, Shuo Yu, Abdul Aziz, Liangtian Wan, Shirui Pan, and Huan Liu.
\newblock Graph learning: A survey.
\newblock \emph{IEEE Transactions on Artificial Intelligence}, 2\penalty0
(2):\penalty0 109--127, 2021.
\bibitem[Xie et~al.(2021)Xie, Wang, Li, Ding, G{\"u}rel, Zhang, Huang, Lin, and
Zhou]{xie2021fives}
Yuexiang Xie, Zhen Wang, Yaliang Li, Bolin Ding, Nezihe~Merve G{\"u}rel,
Ce~Zhang, Minlie Huang, Wei Lin, and Jingren Zhou.
\newblock Fives: Feature interaction via edge search for large-scale tabular
data.
\newblock In \emph{Proceedings of the 27th ACM SIGKDD Conference on Knowledge
Discovery \& Data Mining}, pages 3795--3805, 2021.
\bibitem[Xing et~al.(2019)Xing, Li, Wei, Zhang, Zhan, Zhou, Xue, and
Shi]{xing2019dynamic}
Xiaodan Xing, Qingfeng Li, Hao Wei, Minqing Zhang, Yiqiang Zhan, Xiang~Sean
Zhou, Zhong Xue, and Feng Shi.
\newblock Dynamic spectral graph convolution networks with assistant task
training for early {MCI} diagnosis.
\newblock In \emph{International Conference on Medical Image Computing and
Computer-Assisted Intervention}, pages 639--646. Springer, 2019.
\bibitem[Xu et~al.(2018)Xu, Hu, Leskovec, and Jegelka]{xu2018powerful}
Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka.
\newblock How powerful are graph neural networks?
\newblock In \emph{International Conference on Learning Representations}, 2018.
\bibitem[You et~al.(2020)You, Ma, Ding, Kochenderfer, and
Leskovec]{you2020handling}
Jiaxuan You, Xiaobai Ma, Yi~Ding, Mykel~J Kochenderfer, and Jure Leskovec.
\newblock Handling missing data with graph representation learning.
\newblock \emph{Advances in Neural Information Processing Systems},
33:\penalty0 19075--19087, 2020.
\bibitem[Zhang et~al.(2023)Zhang, Yu, Yan, Liu, Adhikarla, Fu, Chen, Chen,
Zhou, Li, et~al.]{zhang2023biomedgpt}
Kai Zhang, Jun Yu, Zhiling Yan, Yixin Liu, Eashan Adhikarla, Sunyang Fu, Xun
Chen, Chen Chen, Yuyin Zhou, Xiang Li, et~al.
\newblock Biomed{GPT}: A unified and generalist biomedical generative
pre-trained transformer for vision, language, and multimodal tasks.
\newblock \emph{arXiv preprint arXiv:2305.17100}, 2023.
\bibitem[Zheng et~al.(2022)Zheng, Zhu, Liu, Guo, Liu, Yang, and
Zhao]{zheng2022multi}
Shuai Zheng, Zhenfeng Zhu, Zhizhe Liu, Zhenyu Guo, Yang Liu, Yuchen Yang, and
Yao Zhao.
\newblock Multi-modal graph learning for disease prediction.
\newblock \emph{IEEE Transactions on Medical Imaging}, 41\penalty0
(9):\penalty0 2207--2216, 2022.
\bibitem[Zhu et~al.(2006)Zhu, Rosset, Zou, and Hastie]{zhu2006multi}
Ji~Zhu, Saharon Rosset, Hui Zou, and Trevor Hastie.
\newblock Multi-class adaboost.
\newblock \emph{Ann Arbor}, 1001:\penalty0 48109, 2006.
\bibitem[Zhu et~al.(2003)Zhu, Ghahramani, and Lafferty]{zhu2003semi}
Xiaojin Zhu, Zoubin Ghahramani, and John~D Lafferty.
\newblock Semi-supervised learning using gaussian fields and harmonic
functions.
\newblock In \emph{Proceedings of the 20th International conference on Machine
learning (ICML-03)}, pages 912--919, 2003.
\bibitem[Zhu et~al.(2021)Zhu, Xu, Zhang, Liu, Wu, and
Wang]{DBLP:journals/corr/abs-2103-03036}
Yanqiao Zhu, Weizhi Xu, Jinghao Zhang, Qiang Liu, Shu Wu, and Liang Wang.
\newblock Deep graph structure learning for robust representations: {A} survey.
\newblock \emph{CoRR}, abs/2103.03036, 2021.
\newblock URL \url{https://arxiv.org/abs/2103.03036}.
\end{thebibliography}