shakedmanor/gym-members-eda

Fitness Tracker Dataset - Exploratory Data Analysis

Executive Summary: This project presents an Exploratory Data Analysis (EDA) of the "Fitness Tracker Dataset." While officially described as a tool to "provide insights into fitness trends, workout habits, and overall health patterns," what began as a standard investigation into calorie expenditure quickly turned into a data integrity investigation. This notebook documents the discovery of profound synthetic flaws within the dataset, proving mathematically and biologically why it is entirely decoupled from human physiology and unsuitable for machine learning.

1. Dataset Selection & Objective

About the Data: For this assignment, I chose the "Gym Members Exercise Tracking" dataset from Kaggle. It contains physiological and workout records for 1,800 members, including age, weight, heart rate, and various workout types.

Research Question: My primary goal for this Exploratory Data Analysis (EDA) is to answer: "Which factors—such as heart rate, workout duration, or BMI—are the strongest predictors of the calories a person burns during a workout?"

Target Variable: The focus of this analysis is Calories_Burned.

---

2. Dataset Overview & Feature Description

Before diving into the analysis, I examined the raw dataset to understand its composition.

#	Feature	Actual Type	Official Description
1	Calories_Burned	float64	Total calories burned during a session (Target Variable)
2	Age	int64	Age of the individual in years
3	Gender	object	Gender of the individual (Male, Female)
4	Weight (kg)	float64	Weight of the individual in kilograms
5	Height (m)	float64	Height of the individual in meters
6	Max_BPM	float64	Maximum heartbeats per minute recorded
7	Avg_BPM	float64	Average heartbeats per minute during workout
8	Resting_BPM	int64	Resting heartbeats per minute
9	Session_Duration (hours)	float64	Duration of the workout session in hours
10	Workout_Type	object	Type of workout (Cardio, Strength, Yoga, HIIT)
11	Fat_Percentage	float64	Percentage of body fat
12	Water_Intake (liters)	float64	Water intake during or after the workout
13	Workout_Frequency (days/week)	int64	Number of exercise days per week
14	Experience_Level	int64	Fitness experience (Beginner, Intermediate, Advanced)
15	BMI	float64	Body Mass Index ($Weight / Height^2$)

---

3. Data Validation & Discovery of Synthetic Flaws

I performed a series of biological "sanity checks" to verify if the dataset follows real-world logic.

A. Initial Data Audit (Missing Values)

• The Finding: Upon initial inspection of the 1,800 records, 491 missing values were identified across various columns.
• Action: These inconsistencies were flagged for removal during the cleaning phase.

B. BMI Consistency Check

I recalculated the BMI for every member using the standard formula: $$\text{BMI}=\frac{\text{Weight (kg)}}{{\text{Height (m)}}^2}\backslash text\{BMI\}\; =\; \backslash frac\{\backslash text\{Weight\; (kg)\}\}\{\backslash text\{Height\; (m)\}^2\}$$

• Result: In 99% of the rows (1,707 out of 1,800 records), the provided BMI contradicted the mathematical calculation.
• Conclusion: The original BMI feature was generated randomly and was therefore dropped.

C. Heart Rate Logical Sequence

I validated that each record followed the biological sequence: $$\text{Resting\_BPM}\le \text{Avg\_BPM}\le \text{Max\_BPM}\backslash text\{Resting\backslash \_BPM\}\; \backslash le\; \backslash text\{Avg\backslash \_BPM\}\; \backslash le\; \backslash text\{Max\backslash \_BPM\}$$

• Result: 67 instances were found where this sequence was broken (e.g., Avg BPM > Max BPM).
• Action: These rows were filtered out to maintain data integrity.

D. Global Boundary Sanity Checks

I established strict biological boundaries for every key feature. The results were:

• Workout Frequency: All values within 0–7 days/week.
• Fat Percentage: Within logical range (5%–60%).
• Age: Within adult range (18–70 years).
• Height & Weight: Values reflect human dimensions (Height: 1.2m–2.5m, Weight: 35kg–250kg).
• Session Duration: All within logical bounds (0–4 hours).
• Water Intake: Logical consumption levels (0–4 liters).
• Resting, Avg, Max BPM: All values within absolute biological limits.

Conclusion: All feature boundaries are physically and logically realistic.

---

4. Data Wrangling & Cleaning Summary

Action	Description	Affected Features
Fixing Typos	Removed artifacts like \T and converted to numeric.	Max_BPM
Standardization	Cleaned string prefixes like \N and stripped whitespaces.	Workout_Type
Missing Values	Dropped rows with null values to maintain quality.	All Columns
Feature Removal	Dropped invalid BMI features.	BMI, BMI_Category
Row Filtering	Filtered rows violating the biological BPM sequence.	Resting_BPM, Avg_BPM, Max_BPM
Encoding	Mapped Gender to numeric and applied One-Hot Encoding.	Gender, Workout_Type, Experience_Level

Final Cleaned Sample: The dataset is now deeply cleaned and pure, resulting in a final count of 1,301 rows and 19 columns.

5. Exploratory Data Analysis (EDA)

A. Statistical Summary

Before visualizing the data, I generated a statistical summary of the key physiological features. This is where the first major red flag appeared.

Feature	count	mean	std	min	25%	50%	75%	max
Age	1301	34.57	12.25	18	24	33	45	59
Weight (kg)	1301	67.27	19.61	40	52.3	64.8	79.7	129.9
Height (m)	1301	1.73	0.12	1.5	1.66	1.73	1.82	2
Max_BPM	1301	180.97	11.16	160	172	182	191	199
Avg_BPM	1301	145.13	14.96	120	133	144	158	169
Session_Duration (h)	1301	1.38	0.37	0.5	1.14	1.36	1.63	2
Calories_Burned	1301	1031.52	322.64	303	797	1031	1245	1783
Fat_Percentage	1301	23.41	5.90	10	20.2	24.1	27.4	35

• The Anomaly: The average energy expenditure is 1,031.5 kcal per session, with a maximum reaching an absurd 1,783 kcal. Burning over 1,000 calories in a standard workout (which averages just 1.38 hours) is exceptionally high and mathematically improbable for the general population. This "Calorie Paradox" sparked the deeper forensic investigation.

B. Member Demographics & Variance

• Physiological Inconsistency: Height and Weight distributions are nearly identical across genders, proving gender was ignored during generation.

C. Visualizing Feature Independence

The resulting "Uniform Cloud" proves that Age and Weight were generated as independent random variables.

---

6. Correlation Analysis

A. Original Feature Correlation

• The Verdict: Correlation between Calories_Burned and physical drivers is 0.00. The target variable is completely decoupled from metabolic laws.

B. Advanced Diagnostic: BMI vs. Calories

C. The Reality Check: Scientific Benchmarking

I applied the Keytel et al. (2005) formula to calculate expected expenditure:

For Men: $$EE=\frac{-55.0969+(0.6309\times HR)+(0.1988\times W)+(0.2017\times A)}{4.184}\times TEE\; =\; \backslash frac\{-55.0969\; +\; (0.6309\; \backslash times\; HR)\; +\; (0.1988\; \backslash times\; W)\; +\; (0.2017\; \backslash times\; A)\}\{4.184\}\; \backslash times\; T$$

For Women: $$EE=\frac{-20.4022+(0.4472\times HR)-(0.1263\times W)+(0.074\times A)}{4.184}\times TEE\; =\; \backslash frac\{-20.4022\; +\; (0.4472\; \backslash times\; HR)\; -\; (0.1263\; \backslash times\; W)\; +\; (0.074\; \backslash times\; A)\}\{4.184\}\; \backslash times\; T$$

(Where: EE = Energy Expenditure in kcal, HR = Heart Rate in BPM, W = Weight in kg, A = Age in years, and T = Session Duration in minutes).

• The Findings: The dataset artificially inflates calories by 27%.

D. Biological Reality: The Scientific Benchmark

• Result: With scientific calculation, Scientific_Calories correctly correlates with Duration (0.70) and BPM (0.45).

E. Why Workout Type is Meaningless

---

7. Final Conclusion: Research Verdict

The objective of this EDA was to identify which factors drive calorie expenditure in this dataset. Based on my analysis, the answer is: None.

The dataset's target variable is purely synthetic and decoupled from physiology:

1. No Logic: Physical traits (Weight, Age, BMI) show zero statistical relationship with the calorie outcome, contradicting human metabolism.
2. Biological Flaws: The "Uniform Cloud" and the lack of gender-based differences prove the features were generated as independent random numbers.
3. The 27% Gap: Comparing the data to the Keytel Formula exposed a massive inflation bias and confirmed that the underlying heart rate data is equally flawed.

Final Verdict: This dataset is the product of a random number generator, not a biological simulation. It is unsuitable for predictive Machine Learning, as any model trained on this data would produce scientifically invalid results.

Project Metadata

• Full Technical Analysis: Open Google Colab Notebook
• Author: Shaked Manor
• Course: Introduction to Data Science | Reichman University (IDC)
• Video Presentation: