SANTTUcurriculum vitae
03 Sep 2018

A Case Study on Type-1 Diabetes Self-care Survey Responses

Patient Level Analytics Using Self-Organising Maps: A Case Study on Type-1 Diabetes Self-care Survey Responses

Survey questionnaires often with contain missing values. While traditional, model-based methods are commonly used by clinicians, I deployed Self Organizing Maps (SOM) as a means to analyse and visualise survey responses.

Motivations:

  • Clinicians often conduct surveys to better understand their patients. Surveys are heterogeneous with missing values and traditional statistical analysis provide overly simplified solutions.
  • Hence motivation for using machine learning method: SOM.

Advantages of SOM:

  • Mining correlations from the data, 2D visualization and automatic clustering of  features.

Objectives:

  • identify co-morbidities;
  • Link self-care factors that are dependent on each other;
  • Visualise individual patient profiles;

Contributions:

  • Novel use of SOM for visualizing individual patient profiles.
  • Improved understanding for clinicians.

 

Dataset, pre-processing & Demographics

The survey consists of 611 patients’ responses, including 15 questions on self-care factors and 18 for the  co-morbidities  associated with the Type-1 diabetes. The responses for co-morbidities are binary (‘Yes’ or ‘No‘); and for the self-care behaviours, they are one of the following:  1) Never 2) Rarely 3) Sometimes 4) Usually and 5) Always.


Preprocessing:

  • String conversions: Yes (1), No (0 ); Never – Always (1-5).
  • Missing values are imputed using the K-Nearest Neighbour method.

Demographics

Males  and females constitute 45% vs 55%; full time employment , 48.45%; and the lowest percentage of being unemployed is 3.11%. The most common co-morbidity is Retinopathy followed by high BP and high cholesterol. The most common co morbidity found in females is Polycystic ovary syndrome and in males, sexual dysfunction.

 

 

Results

Correlations based on Co-morbidities

Similar outlook of SOM in component planes shows the correlations amongst the features. This is supported by traditional correlation analysis.

 

Patient level analytics: co-morbidities

(a) Output of SOM in U-Matrix  provides an alternative way of  visualizing clustered correlated features in a low dimensional  space

(b). Patient id = ‘111’ is associated with high BP and cholesterol, patient id = ‘211’ with heart disease, patient id = ‘311’ with high BP, high cholesterol and Retinopathy and patient id = ‘611’ with the depression, anxiety and Retinopathy.

Patient level analytics: selfcare factors

(a) Output of SOM in U-Matrix provides an alternative way of visualizing clustered correlated features in a low dimensional space

 

Patient level analytics: co-morbidities

(b). Patient id = ‘111’ is associated with high BP and cholesterol, patient id = ‘211’ with heart disease, patient id = ‘311’ with high BP, high cholesterol and Retinopathy and patient id = ‘611’ with the depression, anxiety and Retinopathy.

 

The component plane shows patient id 77 and his/her dependencies with respect to the self-care factors. The colorbar in this figure shows the user’s rating i.e ‘Never’ corresponding to darker blue, ‘Rarely’ to lighter blue, ‘Sometimes’ to green, ‘Usually’ to yellow and ‘Always’ to orange. It is apparent that patient id = 77 is not wearing a medical alert and rarely keeps food records.

 

Patient level analytics: selfcare factors

The component plane shows patient id  77 and his/her dependencies with  respect to the self-care factors. The colorbar in this figure shows the user’s rating i.e ‘Never’ corresponding to darker blue, ‘Rarely’ to lighter blue, ‘Sometimes’ to green, ‘Usually’ to yellow and ‘Always’ to orange. It is apparent that patient id = 77 is not wearing a medical alert and rarely keeps food records

Data Analytics • health care • machine learning Leave a comment

Leave a Reply

%d bloggers like this: