Gait Event Detection & Pain Anamnesis Classification

Project Overview

This project was conducted as part of the TUM Application Project[IN2328] in association with an industry partner.

The main objective was to develop baseline machine learning models for two related healthcare tasks:

1. Gait Event Detection using IMU sensor time-series data
2. Pain Anamnesis Classification using biomechanical gait-derived features

Two modeling strategies were implemented and compared:

• XGBoost (tree-based gradient boosting)
• LSTM / Deep Neural Networks

The detailed experimental results are documented in the final report.

---

Exploratory Data Analysis (EDA)

Before training the models, exploratory data analysis (EDA) was conducted to understand the statistical properties of the datasets and identify patterns relevant for modeling.

Gait Event Dataset

Dataset Structure

• Total rows: 508,446
• Collection IDs: 2,324

Each collection ID represents one walking session, meaning multiple rows together form a continuous IMU time-series sequence recorded from the same participant.

Each observation includes:

• 3-axis accelerometer signals
• 3-axis gyroscope signals
• timestamp
• foot side (left / right)
• gait event label

Temporal Dependency Analysis

Because the dataset consists of time-series sensor signals, the Autocorrelation Function (ACF) was used to analyze temporal dependencies.

Results:

• Average optimal lag ≈ 19 timesteps

Interpretation:

• Sensor measurements from the previous ~19 timestamps significantly influence the current gait phase.

Based on this result:

• Lag features were created for the XGBoost model
• With 6 IMU channels × 19 lags = 114 lag features

After lag generation:

• Dataset size became approximately 464,590 rows × 116 features

---

Pain Anamnesis Dataset

Dataset Structure

• Total records: 2,603 patients
• Input features: 7 biomechanical gait features
• Targets: 42 pain variables

Each record corresponds to one patient and contains biomechanical features derived from gait analysis using IMU sensors embedded in instrumented insoles.

Feature Variables

The dataset contains seven biomechanical features describing characteristics of a patient's walking pattern and posture:

1. Left movement deviation
   Lateral deviation of the movement trajectory on the left side during walking.

2. Right movement deviation
   Lateral deviation of the movement trajectory on the right side during walking.

3. Resting deviation
   Postural deviation measured while the patient is standing still.

4. Average step length (left)
   Average step length of the left foot during walking.

5. Average step length (right)
   Average step length of the right foot during walking.

6. Average heel strike timing
   Temporal characteristic of the gait cycle describing heel strike timing during walking.

7. Shoe size
   Patient’s shoe size, included as a proxy for foot length which may influence gait mechanics and pressure distribution.

Target Variables

The pain anamnesis dataset contains 42 target variables:

• 18 binary variables → indicate the presence or absence of localized pain (0 / 1)
• 24 ordinal variables → represent pain severity levels on a 0–5 scale

Target Processing for Modeling

To improve the relationship between input features and prediction targets, different target processing strategies were applied depending on the modeling approach.

• Neural Network (LSTM-based model)
  The original target structure was preserved. The model was designed as a multi-task learning framework that simultaneously predicts binary pain presence and ordinal pain severity.

• XGBoost model
  The ordinal targets were converted into binary variables (pain vs. no pain). This simplification reduces the complexity of the classification task and strengthens the statistical association between biomechanical gait features and the target variables.

Feature Distribution

Descriptive statistics were computed for all biomechanical features.

Key observations:

• Most biomechanical features exhibit approximately Gaussian-like distributions, indicating that they can be treated as continuous variables.
• Although shoe size is technically discrete, it spans a sufficiently wide range and was therefore handled as a continuous variable in the analysis.
• Since the feature ranges differ substantially, feature normalization was applied prior to model training.

Therefore:

• Standardization or MinMax scaling was applied before training machine learning models.

Feature–Target Correlation Analysis

Several statistical methods were used to examine relationships between gait features and pain outcomes: To investigate the statistical relationship between biomechanical gait features and pain targets, several dependency measures were used:

• Point-biserial correlation
  Evaluates the relationship between continuous input features and binary pain indicators.

• Distance correlation
  Captures both linear and nonlinear dependencies between variables.

• Mutual information
  Measures the amount of shared information between features and targets without assuming any specific functional relationship.

These analyses indicated that individual feature–target relationships were generally weak, suggesting that pain outcomes are influenced by complex interactions among multiple gait features rather than a single dominant variable.

Results:

• Most feature–target correlations were weak (< 0.3).
• No single biomechanical feature strongly predicts pain.

Interpretation:

• Pain prediction likely depends on complex interactions between multiple gait features rather than a single variable.

Target Dependency Analysis

Relationships between pain variables were analyzed using Cramér’s V.

Results:

• Some pairs of pain variables showed moderate associations (Cramér’s V > 0.4).

Interpretation:

• Certain pain regions tend to co-occur, suggesting shared biomechanical or physiological causes.

---

Task: Gait Event Detection

The goal of this task is to classify gait events from IMU time-series signals. Two different modeling approaches were implemented.

---

XGBoost

Method

Lag-based feature engineering was used to convert IMU time-series into tabular features for XGBoost.

• ACF-based lag selection (average optimal lag ≈ 19 timesteps)
• Lag feature generation (6 IMU channels × 19 lags = 114 lag features)
• MinMax normalization
• Group-aware splitting to avoid leakage across collection IDs (GroupShuffleSplit)

Hyperparameter Optimization (Optuna)

Optuna was used for resource-efficient hyperparameter optimization based on Bayesian optimization (TPE sampler).
The objective was to maximize validation accuracy using a single train/validation split.

The selected hyperparameters control model complexity and generalization:

• learning_rate: 0.07, controlling how much each boosting step contributes to the model and enabling stable learning while reducing overfitting.

• max_depth: 9, limiting the depth of each decision tree while allowing the model to capture nonlinear interactions between IMU sensor features.

• n_estimators: 306, defining the number of boosting rounds used to progressively refine model predictions.

• subsample: 0.52, using a random subset of training samples for each tree to improve robustness and reduce overfitting.

• colsample_bytree: 0.80, randomly sampling a subset of features for each tree to increase model diversity.

• gamma: 0.85, requiring a minimum loss reduction before additional tree splits to regularize the model.

These hyperparameters allow the model to balance model expressiveness and regularization, which is important when learning temporal patterns encoded in lag-based IMU features.

Result

• Test accuracy: ≈ 92%
• Weighted F1-score: ≈ 0.92

Per-class performance highlights:

• Strong performance on frequent classes (e.g., Foot Flat, Toe Off) with scores around ~0.90+
• Lower performance on the rare class Heel Strike:
  • Precision: 0.71
  • Recall: 0.53
  • F1-score: 0.61

Collection-level evaluation:

• ~74% of test collection IDs achieved >90% accuracy (342 / 465 collection IDs)

Distribution shift analysis (Jensen–Shannon Divergence, JSD) was used to quantify the similarity between the predicted and true gait phase distributions:

• JSD excluding Heel Strike: 0.006
• JSD including Heel Strike: 0.12

This result indicates that most distribution mismatch is caused by the Heel Strike phase, while the remaining gait phases are predicted with highly similar distributions.

---

LSTM

Method

Raw IMU sequences were directly used as input to an LSTM network in order to capture temporal dependencies in gait signals.

• Sliding window segmentation (2–3 seconds per window)
• Sequence-based classification
• MinMax normalization applied to all IMU channels

Architecture

The model consists of stacked recurrent layers designed to capture temporal patterns in gait signals.

• 3 stacked LSTM layers
• 128 hidden units per layer
• Dropout applied for regularization
• Fully connected layer with softmax activation for multi-class gait phase prediction

Training

• Optimizer: Adam (learning rate = 0.001)
• Loss function: Cross-entropy loss
• Train / validation / test split: 70 / 15 / 15
• Early stopping used to prevent overfitting

Result

• Overall test accuracy: ~91%

Performance characteristics:

• Strong performance on frequent classes such as Foot Flat and Toe Off
• Lower recall for the rare class Heel Strike, similar to the XGBoost model due to class imbalance

Compared to the XGBoost approach, the LSTM model directly learns temporal dependencies from raw IMU sequences, which allows it to achieve competitive performance without manual lag feature engineering.

---

Task: Pain Anamnesis Classification

Goal: Predict patient pain profile using gait-derived biomechanical features.

---

XGBoost

Method

The XGBoost model was applied to structured biomechanical gait features to predict pain-related targets.

• 7 biomechanical gait features were used as input variables.
• A total of 42 pain-related targets were modeled, and each pain variable was trained as an independent binary classification model.
• To manage the large number of targets efficiently, a custom wrapper class was implemented to train and manage multiple XGBoost models within a unified framework.
• Ordinal pain scores (0–5) were converted into binary indicators (pain vs. no pain) to reduce the complexity of the relationship between input variables and target variables.
• Feature standardization was applied before training.

This approach allows multiple pain prediction tasks to be handled within a single modeling pipeline while keeping each model specialized for its corresponding target variable.

Hyperparameter Optimization

Hyperparameters were optimized using Optuna, which performs Bayesian optimization using the Tree-structured Parzen Estimator (TPE) sampler.
Each pain target was trained as an independent model, and Optuna searched for hyperparameters that maximized validation accuracy.

The optimization also addressed class imbalance, which was common across many pain variables.

Key observations from the optimization results:

• max_depth was typically optimized to ≈ 4, controlling model complexity and preventing overfitting.
• n_estimators centered around ≈ 655 trees, with values ranging from 339 to 799 across different target models.
• learning_rate had a median value of ≈ 0.048, while subsample and colsample_bytree were typically around ≈ 0.892 and ≈ 0.776, respectively, balancing model stability and generalization.
• scale_pos_weight ranged from 1 to 9 with a median of ≈ 3, compensating for imbalanced pain labels.

In addition to hyperparameter tuning, decision threshold optimization was performed.

Instead of using the default probability threshold (0.5), the optimal threshold for each target was selected using Youden’s J statistic:

J = Sensitivity + Specificity − 1

The threshold that maximized Youden’s J on the validation set was selected.
This helped balance false positives and false negatives, which is particularly important in medical classification tasks.

Results

The final XGBoost models achieved the following performance:

• Binary pain prediction accuracy: ~78%
• Average ROC-AUC across targets: ~0.74

Accuracy was used to measure the overall correctness of pain presence predictions across all targets.
Additionally, ROC-AUC was evaluated because many pain labels are imbalanced, and ROC-AUC provides a threshold-independent measure of the model's ability to distinguish between pain and no-pain cases.

Model Interpretation

To interpret the trained models, SHAP (SHapley Additive exPlanations) was used.

• SHAP summary plots were used to visualize global feature importance.
• SHAP dependence plots were used to analyze relationships between biomechanical features and pain predictions.

This analysis helped identify which gait deviations contribute most strongly to specific pain regions and provided interpretable insights into the model behavior.

---

LSTM / Multi-Output Neural Network

Since the pain anamnesis dataset consists of structured biomechanical features rather than time-series sequences, a multi-output neural network was used instead of a recurrent LSTM model.

Input

• 7 biomechanical gait features

Architecture

The neural network was designed as a multi-task learning model with shared feature extraction layers.

• Shared dense trunk: 64 → 32 → 32

Two output branches were used:

Binary pain prediction:

• 18 binary outputs
• Sigmoid activation

Ordinal pain severity prediction:

• 24 ordinal outputs
• Implemented using the CORAL (Cumulative Ordinal Regression with Logistic models) framework

This architecture allows the model to simultaneously learn pain presence and pain severity.

Training

The model was trained using a multi-task loss function combining binary and ordinal objectives.

• Binary Cross Entropy for binary targets
• CORAL-based ordinal regression loss for severity prediction
• Adam optimizer
• Multi-task loss balancing to stabilize training

Results

The multi-task neural network achieved the following performance:

• Binary pain prediction accuracy: ~80%
• Ordinal pain severity prediction: ~1.4 MAE

Compared to single-task models, the multi-task approach improved prediction consistency by allowing the model to jointly learn relationships between pain presence and pain severity across different body regions.

Both the XGBoost and neural network approaches showed limited predictive performance for the pain anamnesis task. This is likely because the direct statistical relationships between the gait-derived variables and pain targets are weak, as indicated by the low pairwise correlations observed during exploratory analysis. Additionally, the relationship between biomechanical gait features and reported pain may be highly complex and multi-factorial, making it difficult for standard models to capture meaningful patterns with the available features. Future work should therefore focus on analyzing model behavior using explainable AI techniques such as SHAP, which can help uncover hidden feature interactions and guide improvements in feature engineering or model design.