Welcome to my new research topic homepage - Biomedical pattern
analysis.
Since March 2011, I have been
working on biomedical pattern/signal analysis, a subject which is at the interface of
knoweldge between pattern recognition and medicine. I work on the
Quality Improvement of Chronic Kidney Disease (QICKD) data set, which
contains the records of about a million patients suffering from
chronic kidney disease (CKD) to various degrees.
The objective
of the study is really to make sense out of the data. One question
that the QICKD investigators are interested in is to estimate the
change in renal function (say g') given the current state of renal
function (g) and the age of a patient (t). Estimating the change in
renal function appears to be harder than we think. This is because each
measurement, known as estimated Glomerular Filteration Rate (eGFR), is
influenced by daily fluctuation such as our body's biological clock
(circadian rhythm), as well as food intake (particularly protein) and
activities performed prior to the measurment being taken. Therefore,
the objective is really to estimate the long term trend amidst the
fluctuation. The conventional method proceeds by calculating the rate
of change given two consecutive eGFR values. I propose to fit a
regression line on the eGFR sequence for each patient. Hence, each patient has a single model - the essence in personalized medicine. The rate of
change of eGFR is then obtained by calculating the first
derivative of the fitted function. The main advantage of this method is
its robustness to instantaneous fluctuation (since we give the expected g)
and the secondly, the rate change of the expected eGFR trend can be
derived analytically and at any given point in time.See the figure below.
The next step consists of sampling g, g' and t and then estimate
p(g'|g,t). This function gives the likelihood of the rate change of eGFR
given the current renal stage (that is the current eGFR value) and the
age of the patient (t). We frame this as a conditional density
estimation problem in which g',g, and t are continuous. The result is a
likelihood graph shown below. The same graph can be represented as a
likelihood table.
Publications:
Poh, N. and S. de Lusignan (2012). "Data-modelling and
visualisation in chronic kidney disease (CKD): a step towards
personalized medicine." Informatics in Primary Care 19(2) [pdf]
Poh,
N. and S. de Lusignan (2011). Modeling Rate of Change in Renal Function
for Individual Patients: A Longitudinal Model Based on Routinely
Collected Data. Neural Information Processing Systems (NIPS)
Personalized Medicine Workshop 2011 (NIPS PM 2011), Sierra Nevada. [pdf] [Download the likelihood table] [spotlight presentation]
