Data reduction and analysis in in-vivo pharmacology models where data is collected for extended periods
February 22, 2009 11:50 pm Pharmacology, ResearchIn models where in-vivo tests are monitored over days and not hours, a tremendous amount of data is generated. How are other researchers handling these large databases and what tools are used to select the subsets of the total data set for analysis?
March 7th, 2009 at 9:16 am
Datacruncher: This is a good question. Data reduction from telemetry studies is a key issue for many, and there are differing approaches that I have seen in the literature, and the specific approach may depend on the volume of data you are trying to analyze, e.g., examine all data collected over 24 hr, or only examine 1 min average of data every hour. In the former, you are trying to analyze as much of the data as possible yet omit extraneous areas of non-usable data, and in the latter, you are analyzing a very small amount of total data (24 min of data per 24 hr period).
Typically, we perform telemetry studies and conduct 24 hour data collection in non-rodent telemetry study, and multiple channels of data are collected, e.g,. ECG, arterial pressure, etc. Our preference is to analyze all the data collected, by breaking down the data into various time bins, e.g., 15 min averages for first 8 hours, then 1 hour averages for the remainder of the study.
To insure that quality waveforms are included in the analysis, there is a need to exclude: areas of mistriggered ECG intervals, areas of lost data (signal drop-out, for instance), areas of artifact due to noise or artifact (emesis, for instance), areas of interference (room entry; dosing), etc.
Depending on your particular software package, there may be options that allow you to scan thru the data visually, to identify problem areas, or you may be able to detect data variability by examing time bins of derived data with high variance, e.g., an animal has a time bin with an average QTc interval of 300 msec with a 250 msec standard deviation around that average, but has a stable heart rate in that same time bin.
In regions with high variance, it is important to assess whether the variance is due to to noise, poor triggering, or the presence of a real signal, like an arrhythmia.
This is one approach, which takes time, because of our preference to analyze all the data. I am aware that other investigators may analyze smaller time bins, e.g., 1 min every 30 or 60 min, which probably requires less time to “crunch”, but a smaller total amount of data is evaluated, and the potential to identify serious effects is lost.
August 19th, 2009 at 5:29 am
I would agree with most of the comments left by HugoV. I would like make sure we separate quantitative and qualitative in this discussion. I have seen studies where the data is collected continuously for up to 72 hours down to 3 minutes an hour and almost everything in between.
Several years ago (circa 2002) I collected 24 h of continuous data with QT positive controls (SPS Poster with astemizole and terfenadine) and then looked at analyzing the data by parsing out different amounts of data down to 1 minute at 6 time points (based on the PK). For the quantitative measures there was very little difference. At that time we overread all the ECG data because we did not feel the software was reliable in marking ECGs consistently. With this technique, more data was not necessarily any better for quantitative measures. We concluded that it was more important to look at less data and use human overread to insure the marks were correctly placed.
In that same study astemizole did induce an arrhythmia that resembled TdP in one of the NHP. There were 6 such events and the longest one lasted ~6 seconds. None of them were found by reviewing the data for a brief period every hour or every half hour. However, the sampling method did work in that there were dramatic changes to the ECG waveform that were found and did give a “signal” that more data should be reviewed.
As was pointed out, even with today’s new software there is the need for human over read and data exclusion. If we take the data we have and retest the various sampling techniques with our new tools, we should be able to better answer the question posed by “Datacruncher”. Is the sensitivity and specificity of the model compromised or perhaps improved by collecting less data and paying closer attention to the data that is collected? Several labs could develop a testing plan and go together to investigate the actual amount of data that should be recorded and the amount that should be analyzed for both quantitative and qualitative analysis of the ECG and hemodynamic data. Any takers?
December 18th, 2009 at 2:33 am
we transfer all data into Matlab, assuming that you can do this (depending on the codes your soft is using, you may need a specific module, to buy or to build up), good luck !