Hi, I’m Jansen Ang
an Aspiring ML Engineer. a Software Engineer. a Tech Community Speaker. a Tech Community Leader. an Open-Source Contributor.

Background: This project is situated at the critical intersection of healthcare and technology, specifically focusing on the use of Machine Learning (ML) to predict diabetes. It is inspired by the alarming statistics from the Philippine Statistics Authority, which ranked diabetes mellitus as the fourth leading cause of death in 2022. Our project is not just a technological venture but also a commitment to the United Nations’ Sustainable Development Goal No. 3, which advocates for universal well-being.

Objectives

1. Develop an Accurate ML Model: To create a reliable model for early diabetes prediction.
2. Feature Set Impact Analysis: To evaluate the influence of different feature sets on the model’s performance.
3. Healthcare Recommendations: To provide urgent care recommendations for those at high risk of developing diabetes.

Scope and Limitations

Our study was defined by specific boundaries and constraints, which will be outlined to understand the extent of our investigation.

Methodology

1. Data Collection and Preprocessing:
   • Source: Kaggle dataset with various features.
   • Cleaning Process: Addressed null values and eliminated duplicates for data integrity and balance.
2. Exploratory Data Analysis (EDA):
   • Findings: Positive correlation between diabetes and factors like age and BMI.
   • Correlation Matrix: Identified key indicators such as General Health, High Blood Pressure, Difficulty in Walking, BMI, and High Cholesterol.
3. Model Building and Evaluation:
   • Focus on Recall: Prioritized high recall to minimize false negatives, using 5-fold cross-validation.
   • Resampling Techniques: Utilized techniques like SMOTE-ENN to enhance model performance, as reflected in improved f-beta scores.
   • Model Selection: Chose Logistic Regression combined with SMOTEENN for its high Recall and f-beta score.
   • Hyperparameter Tuning: Employed Randomized Search and Bayes Search, leading to optimized recall scores.
4. SHAP Analysis:
   • Key Features: High BMI values significantly increased diabetes prediction likelihood; age and general health were also influential.

Recommendations

Based on our findings, we proposed interventions at both government and hospital levels, focusing on factors like BMI, Age, General Health, Physical and Mental Health. These recommendations spanned from public awareness campaigns to specialized health services and infrastructure development.

Model Deployment

• Streamlit Application: Deployed our model locally using Streamlit, featuring a CSV-loading for predictions and a chatbot named DiabetesGPT for user interaction.
• Demonstration: Showcased a fast-paced demo of the model, highlighting its user-friendly interface and functionality.

Conclusion

Our project demonstrates the potential of machine learning in the healthcare domain, particularly for diabetes prediction. By identifying key predictors and providing targeted recommendations, we aim to contribute significantly to early detection and prevention strategies for diabetes, aligning with global health goals.

Project Overview

In our recent data science endeavor, we delved deep into the musical universe of the iconic band Hale, leveraging the Spotify API to analyze their song data. Our research focused on a pivotal question: How can Hale replicate the phenomenal success of their hit, “The Day You Said Goodnight”?

Objective and Tools

Our primary objective was to uncover the secret behind the enduring popularity of “The Day You Said Goodnight” and use

this insight to guide Hale’s future productions. To achieve this, we harnessed the power of Python, Scikit-learn, NumPy, Pandas, and Seaborn, along with the Spotify API, to perform a comprehensive exploration and analysis of Hale’s musical portfolio.

Exploratory Data Analysis (EDA)

Our EDA revealed some intriguing patterns:

• “The Day You Said Goodnight” emerged as the top-streamed song of Hale, outperforming their other tracks in combined streams.
• Both “The Day You Said Goodnight” and “Kung Wala Ka” consistently ranked in the Top 200 Chart from 2017 to 2023, indicating their lasting appeal.
• Interestingly, “The Ballad Of” closely resembled “The Day You Said Goodnight” in audio features but lagged significantly in popularity, securing only a minor fraction of Hale’s total streams.

Methodology and Machine Learning Pipeline

Our approach was systematic, encompassing five critical steps:

1. Data Collection & Processing: Utilizing the Spotify API, we gathered track data and audio features of Hale’s songs.
2. EDA: We conducted a thorough analysis of Hale’s songs, focusing on their popularity and characteristics.
3. Genre Classification & Modeling: We trained various machine learning models to classify songs based on genre and sentiment.
4. Recommender Engine Development: We built a recommender engine based on “The Day You Said Goodnight”, creating three distinct types of recommendations.
5. Streamlit Application: Our findings and recommender system were showcased in a Streamlit app, deployed on Streamlit Cloud.

Our machine learning pipeline involved scaling the audio features using MinMaxScaler, genre and sentiment tagging, and training on models like KNN, SVM, Random Forest, Gradient Boosting, Extra Trees, and XGBoost. The best-performing models were ExtraTrees for sentiment-based classification (63% accuracy) and XGBoost for genre classification (71% accuracy).

Recommender Engine and Recommendations

Our recommender engine, based on cosine similarity of audio features and predicted genre probabilities, suggested songs closely related to selected tracks of Hale. We propose three types of recommendations:

1. Tracks related by audio features.
2. Tracks similar in both audio features and genres (Pop, Alt Rock, Sad OPM).
3. Tracks aligned by audio features and sentiments (Chill, Sad, Happy, Hugot).

The engine offers tailored suggestions, enhancing Hale’s potential to connect with diverse audience preferences.

Strategic Recommendations for Hale

To sustain Hale’s prominence in the OPM scene, we suggest a strategic move: re-recording “The Ballad Of” in collaboration with contemporary artists. This approach aims to revitalize the song, capitalizing on its inherent appeal and audio similarities to their biggest hit.

Further, we recommend leveraging social media trends, particularly on platforms like TikTok, to amplify reach and engagement. Collaborations with prominent Filipino artists such as Juan Karlos, Gloc 9, and Liezel Garcia, along with international stars like Lauv, Noah Cyrus, Maroon 5, and Julia Michaels, could create a powerful synergy, blending styles and fanbases.

Conclusion

Our project stands as a testament to the power of data science in reshaping the music industry. By analyzing Hale’s past successes and aligning them with current trends and predictive models, we offer a roadmap for their continued success. Our Streamlit app, embodying this fusion of art and analytics, paves the way for a data-informed future in music production and promotion.