\"We can identify the time series-like characteristics of our features in the following plot:
\n ![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3246090%2Fe6cd81e3611028cc29984ad2bb0753a0%2F__results___6_0.png?generation=1686323069464357&alt=media\")
\n It is possible to verify that our target vectors consist of boolean values determining the classification of the time series at a given time.
\n ![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F3246090%2F219d76190957ee8caf3ea488b0bef6d8%2F__results___22_0.png?generation=1686323196364340&alt=media\")
I'm here to present our proposed solution for the Parkinson's Freezing of Gait Prediction Challenge…
\nUnlike most of the code notebooks shared, this approach takes into account the fact that our target vector represent the classification of the time-series in a given period of time. Therefore we created a sample XGBoost model, with a multi-class objective, in order to test our hypothesis about using a classifier with time-series features.
\nYou can check the code notebook in here.
\nFirst and foremost, I want to congratulate the winners of this competition for their outstanding work. Your contributions have undoubtedly pushed the boundaries of FOG detection and have paved the way for improved personalized treatment of age-related movement, cognition, and mobility disorders.
\nNevertheless, I wanted to take a moment to thank the competition hosts, the Center for the Study of Movement, Cognition, and Mobility (CMCM), Neurological Institute, Tel Aviv Sourasky Medical Center, for organizing such an interesting and impactful competition. The data provided by the three research groups, along with the support of the Michael J. Fox Foundation for Parkinson's Research, has given us a valuable opportunity to contribute to the evaluation, understanding, and treatment of freezing of gait (FOG) in Parkinson's disease.
\nIn our solution, we aimed to address some of the challenges associated with FOG detection, such as class imbalance and the need for expert intervention. Our primary goal was to develop a machine learning model trained just on data collected from the wearable 3D lower back sensor, which would provide an objective and accurate quantification of FOG episodes.
\nTo achieve this, we made several premises during the development of our notebook code:
\nQuality of TDSCFOG data: We assumed that TDSCFOG, collected in a controlled environment, contained better quality information compared to other datasets. Thus, we focused on predicting DEFOG from TDCSFOG. However, we recognize that training a separate model specifically for DEFOG series might yield better results, and it's worth exploring in future iterations.
Non-time series approach: Although our features have a time-series data structure, we chose to adopt a multiclass/multioutput classification model, due to the fact that our target vectors, represents classifications of the time series data. We believe that this approach can outperform traditional time-series approaches in our objective of FOG event classification using telemetry (IoT sensors) data.
\n\n\"We can identify the time series-like characteristics of our features in the following plot:
\n
\n
\n It is possible to verify that our target vectors consist of boolean values determining the classification of the time series at a given time.
\n
While we are proud of our solution, we acknowledge that there is room for further improvement. Here are some considerations for future iterations:
\nMerge events, tasks, and subjects dataset: By including events, tasks, and subjects information in the training set, we can potentially improve our model's performance and gain a deeper understanding of FOG triggers.
Perform PCA before using KMEANs: Applying Principal Component Analysis (PCA) to the training set before utilizing K-means clustering may enhance the clustering performance and help in feature selection.
Handling imbalanced data: We experimented with various techniques to address class imbalance, but further exploration is needed. Upsampling, downsampling, or advanced methods like SMOTE can be investigated to achieve better results.
Train different models for DEFOG dataset: Given the potential differences in characteristics between DEFOG and other datasets, training separate models dedicated to DEFOG may improve overall performance.
Include time-series related features: Incorporating time-series related features, such as temporal statistics or signal decomposition techniques, might provide additional insights and enhance the model's predictive capabilities.
Experiment with different algorithms and hyperparameter tuning: Our initial solution focused on a specific algorithm, but exploring alternative algorithms and performing thorough hyperparameter tuning could lead to further performance improvements.
In conclusion, our approach represents a humble contribution to the field of FOG detection. We are grateful for the opportunity to participate in this competition and to work alongside such talented individuals within the Kaggle community. Together, we strive to improve the lives of the many people suffering from Parkinson's disease and its debilitating symptoms.
\nYou can check the code notebook in here.
\nBest regards,
\nVladimir Simões da Luz Junior
I'm here to present our proposed solution for the Parkinson's Freezing of Gait Prediction Challenge…
\nUnlike most of the code notebooks shared, this approach takes into account the fact that our target vector represent the classification of the time-series in a given period of time. Therefore we created a sample XGBoost model, with a multi-class objective, in order to test our hypothesis about using a classifier with time-series features.
\nYou can check the code notebook in here.
\nFirst and foremost, I want to congratulate the winners of this competition for their outstanding work. Your contributions have undoubtedly pushed the boundaries of FOG detection and have paved the way for improved personalized treatment of age-related movement, cognition, and mobility disorders.
\nNevertheless, I wanted to take a moment to thank the competition hosts, the Center for the Study of Movement, Cognition, and Mobility (CMCM), Neurological Institute, Tel Aviv Sourasky Medical Center, for organizing such an interesting and impactful competition. The data provided by the three research groups, along with the support of the Michael J. Fox Foundation for Parkinson's Research, has given us a valuable opportunity to contribute to the evaluation, understanding, and treatment of freezing of gait (FOG) in Parkinson's disease.
\nIn our solution, we aimed to address some of the challenges associated with FOG detection, such as class imbalance and the need for expert intervention. Our primary goal was to develop a machine learning model trained just on data collected from the wearable 3D lower back sensor, which would provide an objective and accurate quantification of FOG episodes.
\nTo achieve this, we made several premises during the development of our notebook code:
\nQuality of TDSCFOG data: We assumed that TDSCFOG, collected in a controlled environment, contained better quality information compared to other datasets. Thus, we focused on predicting DEFOG from TDCSFOG. However, we recognize that training a separate model specifically for DEFOG series might yield better results, and it's worth exploring in future iterations.
Non-time series approach: Although our features have a time-series data structure, we chose to adopt a multiclass/multioutput classification model, due to the fact that our target vectors, represents classifications of the time series data. We believe that this approach can outperform traditional time-series approaches in our objective of FOG event classification using telemetry (IoT sensors) data.
\n\n\"We can identify the time series-like characteristics of our features in the following plot:
\n
\n
\n It is possible to verify that our target vectors consist of boolean values determining the classification of the time series at a given time.
\n
While we are proud of our solution, we acknowledge that there is room for further improvement. Here are some considerations for future iterations:
\nMerge events, tasks, and subjects dataset: By including events, tasks, and subjects information in the training set, we can potentially improve our model's performance and gain a deeper understanding of FOG triggers.
Perform PCA before using KMEANs: Applying Principal Component Analysis (PCA) to the training set before utilizing K-means clustering may enhance the clustering performance and help in feature selection.
Handling imbalanced data: We experimented with various techniques to address class imbalance, but further exploration is needed. Upsampling, downsampling, or advanced methods like SMOTE can be investigated to achieve better results.
Train different models for DEFOG dataset: Given the potential differences in characteristics between DEFOG and other datasets, training separate models dedicated to DEFOG may improve overall performance.
Include time-series related features: Incorporating time-series related features, such as temporal statistics or signal decomposition techniques, might provide additional insights and enhance the model's predictive capabilities.
Experiment with different algorithms and hyperparameter tuning: Our initial solution focused on a specific algorithm, but exploring alternative algorithms and performing thorough hyperparameter tuning could lead to further performance improvements.
In conclusion, our approach represents a humble contribution to the field of FOG detection. We are grateful for the opportunity to participate in this competition and to work alongside such talented individuals within the Kaggle community. Together, we strive to improve the lives of the many people suffering from Parkinson's disease and its debilitating symptoms.
\nYou can check the code notebook in here.
\nBest regards,
\nVladimir Simões da Luz Junior