/Healthcare

I don't know why and I don't know how

Primary LanguageJupyter Notebook

Micro-EMA

Framework to select the optimal micro-EMA question that can be asked multiple times a day, this question has a high correlation with the physiological manifestation.

Data description

The data is taken from here

  • processed.csv: each row contains features from 1 minute intervals
  • labeledfeatures.csv: same as processed.csv but it has details of the micro-ema responses as well
  • annotations.csv: 21 rows containing details about the activities and their corresponding timestamps
  • elec.csv: contains the raw data (timestamp, voltage)
  • In labeledfeatures.csv the ema responses are given but they there are no entries for TSST rest and cry rest. The ema responses are the same for all the cry activities

Additional Details

  • Removed the following from In-lab dataset
    • removed P9 because the data doesn't mention event type
    • removed p18 because labeledfeatures.csv not present
  • Removed the following from In-Wild Dataset
    • P101, P107, P111, P112, P113, P114 were removed because of negative Sample voltage.

Setting up the environment

pipenv -r install requirements.txt

Candidate EMA selection:

  • To select the candidate EMAs that have high biserial-correlation with intended stressors.
python ema_stressors_correlation.py
  • The correlation plot of all EMAs with intended stressors is saved at /Plots/Correlation_ema_intended.png

Training the models:

Traditional Machine Learning:

  • Run train_chunk_features.ipynb
  • This notebook creates the following hand-crafted features:
Signal Features
Heart-rate Variability Low freq. energy (0.1-0.2Hz), RMSSD, SDSD
Non-heart-rate variability Mean, Mode, Minimum, Range, $40^{th}$, $60^{th}$, $80^{th}$ Percentile.

  • These features are then used to train SVM, Decision Tree, Gradient Boosting, and AdaBoost.

  • The results after validating the models using LOSO are written to Results/loso_chunks.csv.

Deep Learning:

Using LOSO

  • To run the tensorflow model and validate it using LOSO
python train_chunk_loso.py

Without LOSO

  • To train the model on entire In-Lab data
python train_chunk.py
  • The best model checkpoint will be written to chkpts/v4(2)

Testing the model

  • Run testing_inwild.ipynb.
  • This notebook generates accuracy plots for top 3 EMAs obtained after EMA selection over varying window lengths.
  • All the plots are saved at Plots/In-Wild