|
On this site we collect some information appending our
short article in ERCIM News No. 60, January 2005. This information will be
updated as the research goes on and new results are available.
The analyzed dataWe have data collected during whole night recordings of infants and young children. The total amount of data is 72GB, made up by 1342 recordings. Next figure shows the distribution of these recordings according to the children's age in months.
The above figure shows the distribution against the length of sleeping in hours.
Each recording is made up of some channels, each containing the three main channels:
The StrategyWe split each recording into heartbeats and calculate descriptors of each heartbeat. This feature extraction method seems to be the best way to reduce the huge amount of data to a size easily tractable with various algorithms, while keeping the most possible information. Our previous approach, splitting data into breath-cycle did not prove efficient enough, even in the case of looking for events related to breathing. This first step can already find anomalies, see e.g. the next figure.
The lengths of the heartbeats show at once the periodicity of the sleep and make it possible to determine the sleep phases roughly.
Having traced the heartbeats we can calculate the average muscle-tone for each. This parameter is also very well correlated with the periodicity of the sleep, see on the next figure.
Further parameters are the length, depth and shape of the breaths. The heartbeats, with these numerical parameters, form points in a high dimensional space, a projection of this set is shown next, where a clustering of the points given by the k-means algorithm can also be seen.
Plotting this clustering against time we can show that the clusters found really form connected time intervals.
Contacts: András Lukács László Lukács Data mining in sleep research: http://www.ilab.sztaki.hu/~lacko/sleep
|