Fitness: Diet and Exercise

Design:

Motivation: What does one need to do to achieve their ideal body type? The standard answer is "diet and exercise".

Deliverables:

Slide Deck
Code:
Data Collection
Exploratory Data Analysis
Modeling Initial
Modeling Top
Modeling Script
Champion Modeling Script

• Note that the second modeling script is almost identical except for slight tweaks needed under the champion case

Tools:

• SQL
  • postgress - for setting up the database and SQL query
  • pgadmin4 - for testing sql queries
• Python libraries
  • psycopg2 - Python-PostgreSQL Database Adapter
  • pandas
  • xport (version 3.2.1): for working with .XPT files (SAS output files used by the government). Converted the xport files to csv
  • sklearn
  • seaborn
  • matplotlib

General Workflow:

1. Collect Data
   • Data ReadMe
   • Data primarily from NHANES
   • Stored in SQL database
   • Collected standards for target variable (body fat % standards, and BMI standards)
2. Explore and Clean Data - Exploratory Data Analysis\
   • Revealed Imbalanced Data
   • Determined what Nulls to drop, turn to zero, or impute via KMeans
   • Decide on target variable classification
     • Compared using BMI or Body Fat%.
3. Select Scoring
   1. F1 score (macro and F1_fitness)
4. Modeling Loop:
   1. Features: Select Features
   2. Initial Modeling - Modeling Initial\
      • Mini-loop:
        • Pre-processing: Sampling Methods (OverSampling, SMOTE, ADASYN)
        • Cross-Validation
        • Grid Search for best hyper parameters
      • Compare Models
        • compare models by scoring metric
        • send back to feature selection
   3. Final Modeling - Modeling Top\
      • Select Top Two models and continue evaluating
        • Try a few sampling techniques
        • Try a few different features
      • Graphical options:
        • Confusion Matrix
        • PR-AUC graph
        • Feature Importance / Coefficients
   4. After selecting model, continue with
      • Confusion Matrix
      • Feature Importance / Coefficients

Note: In reality, I went back to the data collection multiple times. I also built some of the modeling tools (such as the mini-loop script 'Sampling.py') before going back to do EDA. Then went back to data collection prior to Final Modeling to

If you are reading this, my lesson learned would be to be more systematic in approach. While I prioritized learning what was interesting and doing what next grabbed my attention, it would have been helpful to have spent more dedicated time early on for data collection. I tried to move too fast at times, and going back to do data over and over again was not helpful. Either do it completely right the first time around, or live with certain amount of data limitations. Definitely do not go back to look at data after every step.

Conclusions:

This is a very difficult problem to solve. I initially thought that with the Data set from NHANES, and all the detailed survey information collected, this would be straightforward. There is detailed information on BMI and body fat percentage (through DEXA scans) and survey information on exercise levels, and specific diet information. Not only diet information like low carb, skip meals, high protein, etc, but also info on what somewhat ate on the day of survey - specific foods and nutrition information.

Unfortunately, a few key issues arose:

• Class imbalance: dataset (fairly representative of the US population) was incredibly biased towards the obese (~73%). The "average/acceptable" level only had 21% representation, with just ~5% identified as fit.
  • To then find anything specific that ties lifestyle to fitness level would be difficult.
• Survey considerations:
  • Lifestyle changes vs. Point-in-time: Those who are obese may be taking action to address their obesity
    • This muddles the data for a point in time survey. Perhaps a certain action/lifestyle is effective in addressing fitness (e.g. calorie restriction). If an obese individual is taking this action, and is getting more fit, but still considered obese; this point in time survey would associate the proper action with this obesity (for this individual case)
  • Respondent's base mentality: This can effect how a respondent chooses to respond
    • Someone who is in shape may be more conscious of what they eat, but not consider themselves as restricting calories. To them, it could be "standard" to not overeat.
• Curse of dimensionality:
  • There are many lifestyle aspects captured in the dataset, and less than 12,000 observations (and just ~600 considered fit).
  • The data becomes very sparse and it is hard to capture meaningful

Because of all those factors (especially as they compound), few insights could be made from the data. The only meaningful point is that more days of vigorous exercise (large increases in breathing or heart rate like running or basketball for at least 10 minutes continuously) is associated with being more fit.

Notes on design: Because most of the variables are categorical, it made the most sense to treat this as a classification problem. However, it can also be considered as a regression problem where body fat percentage is a continuous response variable. However, in designing the problem, it seemed easier to make classes for the response variable and use classification models such as decision tree classifiers.

Other issue:

• Correlated variables:
  • In some cases variables were linearly correlated with others. These should be addressed by removing one or more of the variables. This was not yet done, and would be future work. However, this was taken into account when analyzing feature importance under the kitchen sink approach; and ultimately only one feature was important (vigorous exercise)
    • Total fat vs saturated fat, monosaturated fat, and polystaurated fat). This could be addressed by removing one or more of the variables
    • Days of vigorous exercise vs moderate exercise; recreationally vs for work

Notes on Scoring Rationale:

• Precision: "if you do this, then outcome will definitely happen."

  • Need this high to have credibility with those who follow my advice.

• Recall: "if i promote certain strategies... how many others strategies/scenarios (outliers) am I not capturing?

  • Need this to be high, to minimize outfront critics.

• Actually both are important... I don't want to be missing any viable strategies to achieving a good weight. (So that people can choose which way they want to do things or get mad that they are an exception to my model)

• Thus, to accommodate this balance, selected F1 as the metric.

• Considered ROC-AUC for how good a model generally is, however did not pursue this because it would add additional modeling step to develop the curve.

• Custome Scoring: I also considered and implemented a custom scoring metric. However, given potential overlap and inability to use the scoring to influence the algorithms, ultimately, the F1 score would be easier to interpret and implement.

Other References and Notes:

• Body Fat Norms by American Counsel on Exercise
• Alternative view of bf%:
• BMI categories per CDC
• https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6356293/
• Class Imbalance and precision/recall curves:
  • https://machinelearningmastery.com/roc-curves-and-precision-recall-curves-for-classification-in-python/#:~:text=Generally%2C%20the%20use%20of%20ROC,moderate%20to%20large%20class%20imbalance.
  • https://neptune.ai/blog/f1-score-accuracy-roc-auc-pr-auc
  • https://machinelearningmastery.com/roc-curves-and-precision-recall-curves-for-imbalanced-classification/

To-do list

• Add in more detail/rationale on the modeling results