Comorbidity network for chronic disease: A novel approach to understand type 2 diabetes progression

Chronic diseases management outside expensive hospital settings has become a major target for governments, funders and healthcare service providers. It is well known that chronic diseases such as Type 2 Diabetes (T2D) do not occur in isolation, and has a shared aetiology common to many other diseases and disorders. Diabetes Australia reports that it is associated with a myriad of complications, which affect the feet, eyes, kidneys, and cardiovascular health. For instance, nerve damage in the lower limbs affects around 13% of Australians with diabetes, diabetic retinopathy occurs in over 15% of Australians with diabetes, and diabetes is now the leading cause of end-stage kidney disease. Our research focus is therefore to understand the comorbidity pattern, which in turn can enhance our understanding of the multifactorial risk factors of chronic diseases like Type 2 Diabetes. Our research approach is based on utilising valuable indicators present in pre-existing administrative healthcare data, which are routinely collected but often neglected in health research. One such administrative healthcare data is the hospital admission and discharge data that carries information about diagnoses, which are represented in the form of ICD-10 diagnosis codes. Analysis of diagnoses codes and their relationships helps us construct comorbidity networks which can provide insights that can be used to understand chronic disease progression pattern and comorbidity network at a population level. This understanding can subsequently enable healthcare providers to formulate appropriate preventive health policies targeted to address high-risk chronic conditions.

The research utilises network theory principles applied to administrative healthcare data. Given the high rate of prevalence, we selected Type 2 Diabetes as the exemplar chronic disease. We have developed a research framework to understand and represent the progression of Type 2 diabetes, utilising graph theory and social network analysis techniques. We propose the concept of a 'comorbidity network' that can effectively model chronic disease comorbidities and their transition patterns, thereby representing the chronic disease progression. We further take the attribution effect of the comorbidities into account while generating the network; that is, we not only look at the pattern of disease in chronic disease patients, but also compare the disease pattern with that of non-chronic patients, to understand which comorbidities have a higher influence on the chronic disease pathway. The research framework enables us to construct a baseline comorbidity network for each of the two cohorts. It then compares and merges these two networks into single comorbidity network to discover the comorbidities that are exclusive to diabetic patients. This framework was applied on administrative data drawn from the Australian healthcare context. The overall dataset contained approximately 1.4 million admission records from 0.75 million patients, from which we filtered and sampled the records of 2300 diabetics and 2300 non-diabetic patients. We found significant difference in the health trajectory of diabetic and non-diabetic cohorts. The diabetic cohort exhibited more comorbidity prevalence and denser network properties. For example, in the diabetic cohort, heart and liver-related disorders, cataract etc. were more prevalent. Over time, the prevalence of diseases in the health trajectory of diabetic cohorts were almost double of the prevalence in the non-diabetic cohort, indicating entirely different ways of disease progression.

The paper presents a research framework based on network theory to understand chronic disease progression along with associated comorbidities that manifest over time. The analysis methods provide insights that can enable healthcare providers to develop targeted preventive health management programs to reduce hospital admissions and associated high costs. The baseline comorbidity network has the potential to be used as the basis to develop a chronic disease risk prediction model.