Project Limb Rescue

Cancer patients are at risk of lymphedema, a devastating chronic complication. Our overall aim is to develop a product helping patients to monitor the risk of lymphedema. The product marries wearable devices equipped with photoplethysmography (PPG) sensors and our custom software to detect changes in limb characteristics that are concerning for new-onset, or worsening lymphedema. Limb Rescue Cloud, constitute of Data Base, Web Tool, and Machine Learning modules, establish connections between software, doctors, and data scientists.

Project Limb Rescue Machine Learning Module (LimbRescueML)

LimbRescueML implements four classification algorithms, Support Vector Machine(SVM), Multilayer Perceptron(MLP), Random Forest(RF), and Naive Bayes(NB) to predict lymphedema. Users can compare four classification algorithms, train and evaluate models, and predict with saved models.

LimbRescueML provides dataset generation(Generating Dataset), model training(Training), model evaluation(Evaluation), wave prediction(Prediction), and package installation(Installation) pipelines.

Installation

We recommend using conda to manage packages and dependencies. If it is not installed on your machine, follow the instruction on their website.

• Clone our repository:

git clone git@github.com:Rescue-Heroes/LimbRescueML.git
cd LimbRescueML

• Create enviornment and install dependencies:

conda env create --name LimbResCloud --file environment.yaml
conda activate LimbResCloud

• Install our machine learning package:

python -m pip install -e .

• After installation, remember to activate the enviornment whenever you want to use the package.

conda activate LimbResCloud

Getting Started

Generating Dataset

• Step 1: prepare raw data sets and corresponding annotation file. See Raw Data Preparation for more information.
• Step 2: generate preprocessed datasets for model training and wave prediction. See Data Preprocessing for more information.

Training

Training pipeline provides two functionalities:

• Training model with specific hyperparameters set. Trained model and prediction table will be saved. See Training Model for more information.
• Tuning hyperparameters to find the best model settings. Best hyperparameters set, trained model and prediction table will be saved. See Tuning hyperparameters for more information.

Evaluation

Evaluation pipeline allows users to evaluate saved trained models with test datasets or datasets from labeled new raw data. See Evaluation for more information.

Prediction

Prediction pipeline allows users to make predictions for unlabeled new raw data based on trained model. Raw data preprocess should follow Generating Dataset. See Prediction for more information.

Model Zoo and Baselines

NOTE: The following performance table and confusion matrices are generated based on raw data provided by July 19th, 2021.

We performed five tuning hyperparameters runs and picked the hyperparameters set with the highest validation accuracy as default (default yaml configs) for each algorithm. Then we trained each algorithm with these default hyperparameters set five times and averaged the accuracy and confusion matrix as follows.

Performance table

Accuracy	Train	Test
SVM	0.82	0.67
MLP	0.92	0.58
RF	1.00	0.67
NB	0.62	0.58

Confusion matrix for test set

SVM	Pred. normal	Pred. left	Pred. right
True normal	20.0 (1.00)	0.0 (0.00)	0.0 (0.00)
True left	10.0 (0.50)	10.0 (0.50)	0.0 (0.00)
True right	10.0 (0.50)	0.0 (0.00)	10.0 (0.50)

MLP	Pred. normal	Pred. left	Pred. right
True normal	11.6 (0.58)	0.0 (0.00)	8.4 (0.42)
True left	5.0 (0.25)	14.0 (0.70)	1.0 (0.05)
True right	10.0 (0.50)	0.8 (0.04)	9.2 (0.46)

RF	Pred. normal	Pred. left	Pred. right
True normal	18.8 (0.94)	0.0 (0.00)	0.0 (0.00)
True left	8.4 (0.42)	11.6 (0.58)	0.0 (0.00)
True right	10.4 (0.52)	0.0 (0.00)	9.6 (0.48)

NB	Pred. normal	Pred. left	Pred. right
True normal	20.0 (1.00)	0.0 (0.00)	0.0 (0.00)
True left	15.0 (0.75)	5.0 (0.25)	0.0 (0.00)
True right	10.0 (0.50)	0.0 (0.00)	10.0 (0.50)

People

Sponsors: Carlo Contreras, Lynne Brophy

Technical Team:

• Tai-Yu Pan package manager, designed all pipelines
• Mengdi Fan implemented model training and evaluation pipeline, generated documentation
• Rithvich Ramesh tested the Gaussian Naive Bayes algorithm
• Browy Li tested the Random Forest algorithm