Forecasting Potential Diabetes Complications

发布时间：2026年11月4日

Introduction

Diabetes complications often afflict diabetes patients seriously: over 68% of  diabetes-related mortality is caused by diabetes complications. In this paper, we study the problem of automatically diagnosing diabetes complications from patients' lab test results. The objective problem has two main challenges: 1) feature sparseness: a patient only takes 1.26% lab tests on average, and 65.5% types of lab tests are taken by less than 10 patients; 2) knowledge skewness: it lacks comprehensive detailed domain knowledge of association between diabetes complications and lab tests.  To address these challenges, we propose a novel probabilistic model called Sparse Factor Graph Model (SparseFGM). SparseFGM projects sparse features onto a lower-dimensional latent space, which alleviates the problem of sparseness. SparseFGM is also able to capture the associations between complications and lab tests, which help handle the knowledge skewness. We evaluate the proposed model on a large collections of real medical records. SparseFGM outperforms (+20% by F1) baselines significantly and gives detailed associations between diabetes complications and lab tests. 

 

Problem Formulation. We first introduce the observed lab tests, which are performed on patients to evaluate the patient's  health condition. Formally, we define a lab test record of a patient as follows: 
 
 
Generally, our goal is to diagnose diabetes complications from a set of lab test records of patients. 
 
 

Codes

 

We develop a Sparse Factor Graph Model (SparseFGM) The basic idea here is that we first project instances from the original sparse space onto a lower dimensional latent space. The latent space is used to capture the correlation between different lab tests. In this way, the model alleviates the feature sparseness problem. Moreover, the graphical structure of the SparseFGM model is designed to model the correlation between different labels. 

Figure 1 shows a simple example of an SparseFGM. Instances in the input layer (represented as 9-dimensional vector) are sparse. The model projects all instances onto a 5-dimensional latent space, which alleviates the sparseness issue. Further the factor function (indicated as black rectangles) between labels is defined to model the dependencies between instances. Please notice that SparseFGM can not only be applied in diabetes complication forecasting problem but also other similar scenarios. 
 
Figure 1. An illustration of the proposed model. 
 
C++ Codes Download [available soon]

Experiments

 

Data Set. We use a collection of real medical records from a famous geriatric hospital. The dataset spans over one year, containing 181,933 medical records corresponding to 35,525 unique patients and 1,9451,945 kinds of lab tests in total. On average each clinical record contains 24.4324.43 different lab tests (1.25%1.25\% of all lab tests), which indicates the feature sparsity problem is serious for this problem.
 
Results. Based on SparseFGM,  we can discover the underlying patterns between diabetes complications and lab tests at micro-,  meso-, and macro-level. 
At micro-level, we calculate how each lab test is associated with different complications. Figure 2 shows the results obtained in our experiment to demonstrate the association between ten complications and nine parameters of routine urine analysis. The color in each square represents the association strength, discovered by our proposed model, between each urine test and complication. For example, protein (PRO) has strong association with hypertension (HTN) as chronic hypertension causes kidney damage, which in turn leads to the appearance of protein in the urine. 
 
Figure 2. Association between 10 complications and 9 parameters of routine urine analysis. 
 
We demonstrate the results of the meso-level analysis in detail. Generally, we aim to study how different groups of lab tests associated with diabetes complications. The grouping information could be extracted from the proposed model naturally as SparseFGM clusters features into K groups implicitly.  Table 1 presents the top 3 groups of lab tests which are mostly predictive to coronary heart disease (CHD) and hyperlipidemia (HPL) respectively. For each group, top lab tests with the highest probability belonging to a particular group is listed.
 
Table 1. Meso-level analysis.
 
At macro-level, we study the different strength of diabetes complications forecasting depending on lab test results. We say a complication is more diagnosable from lab test results if the proposed model is able to target more positive instances correctly among all positive instances. Thus the strength can be estimated by the measurement recall. Figure 3 shows the results, from which we see that hyperlipidemia (HPL) can  be diagnosed based on lab tests precisely, while depression (depr.) is usually recognized from psychological investigation instead of physiological lab tests such as blood test. 
 
Figure 3. Macro-level analysis.

References

Yang Yang, Walter Luyten, Lu Liu, Marie-Francine Moens, Jie Tang, and Juanzi Li. Forecasting Potential Diabetes Complications. In AAAI'14, 2014.