On this notebook I will try to use several unsupervised methods of learning to find some patterns or other useful information which can help to predict FoG-events. \n
\n I will use and describe next methods:\n
stumpy \n \n Hope this one will useful and helpful to you.\n
\n Enjoy!
\nThe data series include three datasets, collected under distinct circumstances:\n\n- The tDCS FOG (tdcsfog) dataset, comprising data series collected in the lab, as subjects completed a FOG-provoking protocol.\n- The DeFOG (defog) dataset, comprising data series collected in the subject's home, as subjects completed a FOG-provoking protocol\n- The Daily Living (daily) dataset, comprising one week of continuous 24/7 recordings from sixty-five subjects. Forty-five subjects exhibit FOG symptoms and also have series in the defog dataset, while the other twenty subjects do not exhibit FOG symptoms and do not have series elsewhere in the data. \n
\n
I will concentrate on daily dataset only\n
Here we will get a short view on our subjects and other metadata
\n My suggestion is that huge files are from daily series and smaller one are from defog\n
\n Let's check it out\n
\n Well, maybe i was right and all files with size above 500 MB has they time series at defog\n
\n Anyway, we will take one file with huge size and one file with smaller one \n
\n To save your time I will research results of processing only in one accelerometer\n
\n\n Here we can see a raw data plot of AccV values. Next I will try to make this plot better and find a valuable info in it\n
\n Also we can note the X-axis units - it's a seconds for now, but we can get time stamp of record begining in metadata to create a time axis with hours and minutes. Save this idea for now\n Next step is plotting rolling mean to see a main trend in this series\n
\n First we transform a time in metadata to datetime64\n
\n Next we'll find our subject in this data and get his timestamp of begining\n
\n\n Our subject has started his recording session in 12:00. Now we can add a time in seconds to his start time and get a new axis units\n
\n Of course this date is fake, don't watch on this please\n
\n And now we can get a index in datetime64\n
\n That's it, now we can turn to the main part of the show\n
\n\n Rolling mean is a mean of a certain number of previous periods in a time series\n
\n It helps us to see the main line of trend in series or to slightly smoothing the peaks\n
\n Even such a simple method as rolling mean can be useful in prediction of FoG-events\n
\n Note: Here I will create a slice from such a big dataframe, but on local machine you can skip it\n
\n\n So rolling mean give us some more useful view on our series, but there a scale problem: this plot is too small to understand\n
\n Also the window size gives a great impact to plotting results, so this value should be tuned with accuracy\n
\n Note: rolling mean cutting all the peaks in series\n
\n Decomposition - simpliest method to get the main patterns in time series\n
\n Here we can see nexts results of decompostion:\n
\n STUMPY is a powerful and scalable Python library that efficiently computes something called the matrix profile -
\n a vector that stores the z-normalized Euclidean distance between any subsequence within a time series and its nearest neighbor.\n
\n The most interesting possibilities of stumpy for FoG-events prediction are:\n
\n We can do many interesting things with stumpy, first of all we will find best_motif:\n
\n it is the one where the time series is the smallest (or reach the minimum in other words)\n
\n All this gray rectangles are global minimums in our series, possibly it's 'calm' values of accelerometer or another valuable feature\n
\n My suggestions is that window size can make this picture cleaner\n
\n The semantic segmentation goal is find a k areas in whole time series\n
\n Let's take a look\n
\n The result of this method is creating 4 areas (as regimes equals 4) in time series.\n
\n As we can see here a border between our regimes are too narrow and to use this method we should take a smaller time series part\n
\n
\n Time series chains are patterns that evolve and drift over time\n
\n\n Heree we see several parts of time series\n
\n\nShapelets are subsequences that can be used to represent a class. Matrix profiles make it possible to identify these shapelets\n
","metadata":{}},{"cell_type":"code","source":"wavelets = stumpy.stump(subject_slice_stumpy['AccV'], m)[:, 0].astype(float)\nfig, ax = plt.subplots(figsize=(24, 8))\nsns.lineplot(data = wavelets, \n legend = True,\n ax = ax);\nsns.lineplot(y = subject_slice_stumpy['AccV'],\n x = subject_slice_stumpy.reset_index(drop = True).index,\n legend = True,\n ax = ax);\n#ax.set_ylim(-1,0.5)\nax.legend(['Matrix profile', 'Raw series']);","metadata":{"execution":{"iopub.status.busy":"2023-06-05T09:41:40.723188Z","iopub.execute_input":"2023-06-05T09:41:40.723647Z","iopub.status.idle":"2023-06-05T09:41:45.532068Z","shell.execute_reply.started":"2023-06-05T09:41:40.723612Z","shell.execute_reply":"2023-06-05T09:41:45.530639Z"},"trusted":true},"execution_count":null,"outputs":[]},{"cell_type":"markdown","source":"\nShapelets searching is most difficult to me and this is only demonstration of possibilities of stumpy.\n
\n We can go further and build a Shapelet Based Model to find a classes in time series(e.g. FoG-events)\n
\nThere still work to do and now extraction of valuavle information is almost impossible.\n
\n However, matrix profile from stumpy can be illustrative and even now we can see some patterns of events \n
\nI gave a short description to several tools of time series analysis:\n