The Study

To test (or maybe rather get some first impression), whether it is possible to predict subjective experiences of fatigue, for 35 days we (1) took two muscle measurements (see protocol) - one in the morning after waking up, one in the evening before going to bed and (2) filled out a questionnaire (see survey).

We had three main outcomes of interest:

For the remainder you can pick a person whose data you want to explore:

The below plot shows the three main outcomes of interest over time: overall fatigue, muscle soreness and muscle tiredness. Everyday we retrospectively assessed or subjective experience of those. Values range from 1 (= very low) to 7 (= very high). It further shows the levels of some other variables in relation to their respective maximum, e.g. if step count is dark purple, the number of steps on that days was close to the maximum number of daily steps measured during the study period.

Fraction of maximum value drawing
Fatigue levels drawing

Predictions based on Survey Data

Are there patterns in our wellbeing that we aren't conciously aware off?

In a first step, we explored whether we can predict subjective experience of fatigue, muscle soreness and muscle tiredness for the next day based on information collected through the questionnaire the day before. We tested several machine learning models and compared them to (1) our own prediction of how we would feel the next day and (2) always predicting the mean level.

Predictions for

Models

Model Selected:

Overall Fatigue

Muscle Soreness

Muscle Tiredness

Results

Well... it seems like fatigue levels and behavior are not too much indicative of the next days' fatigue levels.

Amanda

Generally speaking, Amanda's predictions of her muslce soreness & tiredness levels for the next days, were slightly worse than just always predicting the mean level observed within the previous month. For the overall fatigue level her own predictions were slightly better - mostly Amanda is a bit too pessimistic.
When it comes to the different models, in terms of absolute error, all of them perform worse than just always predicting the mean.

Johanna

Generally speaking, Johanna's own predictions for overall fatigue and muscle tiredness were worse than always just predicting the mean. For muscle soreness, her predictions were fairly accurate - at least better than all machine learning modelsor simply choosing the mean.
When it comes to the different models, standard random forest performed best for overall fatigue and muscle tiredness. However, it only slightly outperformed always predicting the mean value. While it was quite a bit better than Johanna's own predictions, looking at the SHAP values tells us that Johanna's prediction for the next day nonetheless was the most important variable for predicting. For muscle tiredness also scoring high on the item 'I felt fit' actually increased the likelihood of high muscle tiredness the next day, indicating that feeling fit and potentially moving a lot might have negative effects on muscular fitness the day after. For muscle soreness, lasso regression performed best - looking at the coefficients tells us that, somewhat non-surprisingly, doing HITT exercise was the most relevant predictor of muscle soreness.

Note: We also tested a standard linear regression model but it performed even worse than any of the more complex models, thus results are not includeded here. However, it is included in the code .

Predictions based on muscle data

In addition to the daily survey data, we also collected sensor data about muscle activity twice a day, using an EMG & MMG sensor kit , following a strict protocol . A large body of literature has been trying to use these types of data to assess and/or predict (muscle) fatigue (see here for a great literature review or here for our summary) - with varying success.

Based on the existing literature, we developed a protocol that we hoped would allow us to gather data with high enough quality to be of use while still being short and easy to follow such that we could integrate it in our daily schedule.

Pre-Processing

In a first step, the EMG and MMG data was pre-processed. In this case we (1) substracted the mean and applied a bandpass filter with a lower bound of 50 and an upper bound 450 Hz, (2) rectified the signal and (3) performed amplitude normalization. The last plot also shows the starting points of the different phases in the protocol , where 'M' stands for muscle contraction.
In the next steps, we extracted features, i.e. transformations of the signal based on the features that have been reported to be predictive of muscular fatigue in previous studies. Those were always extracted for sub-signals of a length of 250ms (as recommended by previous literature), i.e. each measurement recording was 'chopped up' into many smaller instances.

Predictions

We used those to train two different types of models: a neural network and a random forest model. Both were trained to predict for every sub-sequence and the final result, i.e. the prediction of fatigue for a particular day as obtained by averaging over all the predictions for that day based on each subsequence. Apart from the features extracte from the signal the models were also given indicators for which interval of the recording a subsequence belonged to (e.g. wallsit) and whether the recording had been created in the morning or evening.

Results

Overall, the muscle data based models performed as badly as the models based on the survey data and our own predictions. There are existing studies using EMG and MMG data to predict (muscular fatigue), however, all of the ones we found were focused on the particular moment of data collection, i.e. predicting e.g. the onset of fatigue during an exercise or contraction based on real-time EMG measurements. This is contrary to our exploration where focused on predicting a general 'state' of muscle impairment and wellbeing. Thus, one conclusion that we might draw is that the electrical activity in the muscle's response to a neural stimuli is not indicative of it's more general state. Alternatively, it could also be the case that it is indeed indicative but our subjective experience of (muscle) fatigue is not related to the actual, physical fatigue state. However, various studies focusing on measuring muscle fatigue have successfully used questionnaires to determine fatigue thresholds. Consequently, we consider the first explanation as more likely.

Predictions for

Models

Overall Fatigue

Muscle Soreness

Muscle Tiredness

Discussion

Of course there could be different reasons why we aren't able to predict next day fatigue very well:

Nonetheless, we both learned a lot - about medicine, muscles, our own bodies, data privacy, interactive design and - last but not least - art-science collaboration. However, if you are reading this and have ideas for improvement or think you have spotted some 'mistakes' in the way we approached things (feel free to also take a look at our code which is open source), we would be excited to hear and learn from you. You can just send us an email.