2

I have a data set containing a daily sensor data measurements recorded from 20 participants for 60 days (baseline data).

I am trying to develop methods for predicting/estimating decline in long-term monitoring studies, i.e. can measurement of a parameter on a daily basis be used to detect/predict decline (i.e. significant change) by identification of negative trends or abberant measurements. I hope to be able to generate statistical thresholds to allow identification of decline by examining the trends (using some combination of sensor derived parameters and referencing these to baseline clinical data.

What is the most appropriate method to define a threshold for a significant change or decline for each participant and how do I best predict/detect negative trending or aberrant behaviour in unseen data?

I wondered if I could get some opinions on a good approach? (note I have been using ICC, ANOVA as well as examining std. dev. of the baseline but have not found any of these approaches particularly useful)

BGreene
  • 3,045
  • 4
  • 16
  • 33
  • Perhaps you could tell us a bit more to help us think about your question. Do you have any idea what amount of decline is important to detect? If so, what is that amount? Is 60 days a long time period for your purposes or a short amount of time? Are you planning to repeat this study or is this data set all you plan to gather? Thanks. – Joel W. Sep 16 '12 at 02:09
  • Just edited to better explain the question. – BGreene Sep 20 '12 at 10:19
  • 1
    Is this restatement correct: You have one measure per day for 60 days and want to predict declines later in the 60 days based on measures earlier in the 60 days. (I do not know what you mean by aberrant measures.) You may want to get change measures, perhaps 59 day-to-day change measures. Then you can see if these change measures predict later (larger) changes that are of interest to you. If variance is large, you might look at changes in week to week averages for each participant. Does that make any sense in light of your specific project? – Joel W. Sep 20 '12 at 13:32
  • I am measuring each patient using a sensor once per day. There are multiple measures derived from the sensor per patient per day – BGreene Sep 20 '12 at 14:55
  • How do you define a decline? Is it a change in one measure, a change in an average of several or all the measures, or some clinical judgment? – Joel W. Sep 20 '12 at 17:14
  • Editing above to better explain – BGreene Sep 21 '12 at 16:01
  • Is this correct: you will have one clinical measure of functional decline, separate from the measures you collect on each of the 60 days. Will that clinical measure be all or none, categorical, or continuous? If categorical, will the categories form a continuum? If continuous, do you consider the data to be interval level? Or will you have more than one clinical measure of decline? – Joel W. Sep 21 '12 at 19:04
  • I have a number of clinical measures captured before the trial started, some categorical, some continuous. Daily data is then recorded once the trial started. Unfortunately there will not be a post trial clinical assessment. – BGreene Sep 23 '12 at 09:26
  • Are all the measures of functional decline collected at the end of the 60 days, or are you looking for trends in functional decline during the 60 days? In any case, what does that measure of functional decline look like? Do you have one measure of functional decline? Is it a measurement or a clinical judgement? If the second, will that clinical measure be all or none, categorical, or continuous? If categorical, will the categories form a continuum? If continuous, do you consider the data to be interval level? Or will you have more than one clinical measure of decline? – Joel W. Sep 23 '12 at 11:43
  • Some more edits above. All clinical measures obtained beforehand. There are a collection of these messures: some clinical scales with continous scores, some are dichotomous variables. To start I will relate the sensor data to the continuous clinical scales (using them as surrogate measures of decline) taken at baseline but will also seek associations by examining the dichotomous variables – BGreene Sep 24 '12 at 14:52
  • Your recent edits provide some more information about your project, but I still do not understand what you are trying to predict. Without knowing that, it is impossible to make any suggestions about how you might analyze you data. – Joel W. Sep 24 '12 at 21:03
  • I am hoping to find some way to dichotomize patients into declined/non-decline groups, i.e. can I search for patterns in the time series of participants that had a decline over the period compared to those how didn’t. Before the trial participants receive a clinical evaluation which includes measures of functional balance etc but we will not have a post-trial assessment – BGreene Sep 25 '12 at 11:19
  • Oh, perhaps now I understand. Is this correct: You have two groups of patients (decline and not decline) and numeric data on each patient. You want to see how well you can predict group membership based on your numeric data. If this understanding is correct, this is a standard problem and many on this forum should be able to help you. If this understanding is not correct, what is it you want to predict? – Joel W. Sep 25 '12 at 16:14
  • Thats not quite it, as the funtional measures are taken beforehand. I want a way to detect if they have decline *since* their assessment based on the trends of their sensor data. Thats why I'm speaking of defining a baseline (based on sensor data but perhaps references to clinical assessment) then detecting deviations from that baseline as being indicative of 'decline' (decline is not clearly defined in this case). Apologies for the ambiguity – BGreene Sep 25 '12 at 17:10
  • Do you have two groups: declined and not-declined? Or do you have one group and you want to detect unusual changes in each individual's data that you collected over the 60 days? (The changes may be declines or improvements?) Or is it something else? – Joel W. Sep 25 '12 at 20:05
  • I have one group and want to detect unusual changes over the 60 days. The study is not fully complete as yet hence some of the ambiguity – BGreene Sep 25 '12 at 20:14
  • So, there seem to be three parts to your question. First, you have 20 people of known health and you will follow them for 60 days. You want to know if any of them show unusual changes in their data over those 60 days. Is that correct? Second, you want to "detect/predict functional decline by identification of negative trends or aberrant measurements." But you also say you will collect clinical data only before the 60 days start. How will you predict decline if you do not collect clinical data after the 60 days? What will be your definition of decline? – Joel W. Sep 27 '12 at 01:24
  • As to the third part to your question, what do you mean when you say you want to use "some combination of sensor derived parameters and" reference "these to baseline ... pre-trial clinical assessment ... and a daily health questionaire." What is the goal of this third part? – Joel W. Sep 27 '12 at 01:24
  • probably best to discuss this in chat. As you say, there are several questions here but I think most can be underpinned by the same analysis. To answer your comments above: yes to the first one. I am hoping the 3rd part can help answerr the quaestion of how to define decline. – BGreene Sep 27 '12 at 09:49
  • Chat sounds ok. I've not used it and do not know how to set up a chat. If you know, let me know when you would like to chat. My schedule is open today and tomorrow, then cluttered for a week. – Joel W. Sep 27 '12 at 17:07

0 Answers0