Hi everyone,
I am trying to predict some diseases using health check data.
The attributes are Age, BMI, Waist, Systolic, Diastolic blood pressure, HDL, LDL ... etc. I have to calculate distances (for kNN ...) or similarities (for Rough Sets ...) among instances. Certainly, we can use Euclidean distance and distance based similarity as well. In fuzzy approach it also possible to calculate similarity between instance "x" and "y" in attribute "a" by similarity_a(x,y) = 1 - |a(x) - a(y)|/(a_max - a_min) and combine with other attributes by a t-norm. However, those calculations are for arbitrary attributes which are mutually independent.
So for health check data, is there any considerable way to calculate distance/similarity?
Thanks in advances,
Hi,
You could use the Heterogeneous Value Difference Metric (HVDM). In this way, you could work simmilarities for numerical and nominal attributes. In the case of medical attributes such as HDL, blood pressure I would try an experiment taking as a basis the "normal" values and calculate the deviation from them. Hope you find this useful!
Dear Dr. Jesus A. Gonzalez,
Thank you very much for the suggestion. I will find and learn about HVDM.
By the way, my concern is how we can calculate the distance in health check data where some attributes may be dependent, for example, the association between HDL and LDL, systolic and diastolic pressure.
Hi,
You can combine Distance Measures(Like ReliefF) and Information Measures(Like Pearson correlation coefficient) .
For more information, see the following article:
"Feature Selection for Classification" Authors M. Dash, H. Liu
Hi, Of course, the following article can help you
"Tuning ReliefF for Genome-Wide Genetic Analysis"
- Badges
- Science topic
- Similar topics
- Social Psychology
- Attitude