wearables

Keywords: mobility data, gestational age

Background

Biological

Methodological

Time-series classification

Models that use each time-point as a feature cannot take advantage of the information within the sequential order of the data (e.g., KRR on aligned & fixed-width sequences). To incorporate this information, RNNs and CNNs (e.g., InceptionTime) are commonly used model dynamics (see sktime-dl).

For small data and non-DL models, implemented in sktime:

• kNN w/dynamic time warping
  • use DTW as a distance metric for a k-NN classifier
  • advantage: good benchmark because of its simplicity
  • disadvantage: performs poorly if noise in series is larger than subtle differences in shape
• TimeSeriesForest
  • RF for time series by splitting series into random sub-sets with random start/length/stop points, extracting primitive values from each interval, and training a RF on those extracted features per series
  • advantage: can be interpreted
• BOSS/cBOSS
  • dictionary-based that uses a sliding window to transform series into a "word" of specified length and of specified number of letters
  • apply Symbolic Fourier Approximation (SFA) transformation by calculating FT on each window, discretizing first l Fourier terms and concatenating to form a "word" using Multiple Coefficient Binning (MCB), a supervised method that bins continuous time-series into a sequence of letters. As the window slides, a count of each word's frequency is recorded in a dictionary, then converted into a histogram. Then, any classifier can be trained on these word histograms and the words can be interpretted (BOSS usess nearest-neighbor classifier and gridsearch to select length of words, number of letters, FT normalization; cBOSS is a contractable version, randomly sampling from parameter space w/o replacement and retaining a fixed-number of classifiers, since BOSS gridsearch has large memory and time requirements w/o much loss in performance)
  • disadvantage: large time and memory requirements and grid-search procedure; also, classifiers may be too simplistic (kNN)
  • intuition: dictionary-based classifiers perform well when you can discriminate using the frequency of a particular pattern
• RISE
  • single random interval from time-series, and instead of extracted primitive values, and relies on concatenated spectral features of the interval for RF classifier, e.g., fitted auto-regressive coefs., est. autocorr coefs., and power spectrum coefs. (i.e., series-to-series feature extraction transformers)
• Shaplet Transform Classifier
  • enumerate intervals, to extract sub-shapes of time-series, which are useful to detect "phase-independent localized similarity between series". After identifying the top-k shapelets (usually evaluated by information gain, and strongest non-overlapping shapelets are retained) to become k features, any classifier can be trained (in the past, weighted ensemble classifier, and Rotation Forest, which is a tree-based ensemble on PCA of the input features, have been used... for continuous features, rotation forest works well but is not available in python)
  • advantage: shapelet features can be used for interpretability... presence of shapelets may indicate something discriminative
  • disadvantage: enumeration of all intervals can be costly, so sktime randomly searches for shapelets, potentially missing something important
  • intuition: shapelet-based classifiers perform well when the "best" feature is the presence or absence of a phase-independent pattern in a series
• HIVE-COTE or ROCKET
  • the HIVE-COTE ensemble classifier can combine all of the above (Hierarchical Vote Collective of Transformation-based Ensembles), e.g., per time-series, feed to Shaplet Transform, TimeSeriesForest, RISE, and cBOSS, then take a weighted average of predictions with a control unit that assigns weights to classifiers by their quality.
  • ROCKET is a simple linear classifier based on random convolutional kernels such as random length, weight, dilation, padding, etc.

Feature extraction of time series can be global or local (sliding windows, or bins), transforming the time series into primitive values (mean, sd, etc.) or other series (FT, series of auto-regression coefs.).

REF: Alexandra Amidon blog on sktime, see here. sktime can be found here. Loning et al. sktime, NeurIPS, 2019

NOTE: the primary goal with wearables is to convert these models for use in time-series regression tasks, where a time-series is used to predict an output value in order to do substite testing in healthcare (e.g., AppleWatch using activity monitoring to estimate a patients six-minute walk test value). However, the above are useful in predicting some other interesting secondary-outcomes of interest in order to gauge how much useful information the time-series contains.

Time-series regression

Data

Mobility or activity data

• https://www.kaggle.com/shambhavimalik/activity-data
• https://wwwn.cdc.gov/nchs/nhanes/search/datapage.aspx?Component=Examination&CycleBeginYear=2003 (see dataset description in https://www.nature.com/articles/s41598-019-46850-0#Sec2 for more details)

Future directions

Development roadmap

1. Incorporate more DL models for ECG based data, use fewer and fewer leads (starting from 12-lead ECG data)

References

1. On the broad utility of activity monitoring: https://doi.org/10.1126/science.abc5096 and how to model with atypical architectures (i.e., GNNs): https://arxiv.org/pdf/2007.03113.pdf; using Google's Community Mobility Reports (https://www.google.com/covid19/mobility/) and The Times COVID-19 database (https://github.com/nytimes/covid-19-data)