Kairos
Problem & Motivation
With the growing complexity of personal healthcare regimens, including the use of multiple medications and supplements, individuals face a significant risk of adverse drug interactions influenced by their lifestyle habits and existing conditions. This app addresses the critical need for a reliable tool that predicts the consequences of concurrent substance use in real-time. Our solution, leveraging a user-friendly OCR and chatbot interface, allows individuals to input their medical profile details to receive immediate feedback on potential reactions of certain drugs. Assumptions include users' willingness to accurately report their health data and the app's ability to process diverse pharmaceutical and dietary substances effectively. This solution targets a gap in personalized, proactive healthcare management, marking a substantial opportunity for impact and user engagement.
Data Source & Data Science Approach
Datasets
DrugBank (Approval Process Complete)
-Drug-drug interactions: Information on known interactions between different drugs, including the severity of the interaction (e.g., no reaction, mild, moderate, or severe).
-Drug properties: Chemical and pharmacological properties of drugs, such as molecular weight, structure, and mechanism of action.
-Drug indications: Therapeutic uses for each drug for context of potential interactions.
-Drug side effects: Known adverse reactions associated with each drug.
ChEMBL
-Drug targets: Information on the biological targets (e.g., proteins, enzymes) that each drug interacts with, which can help predict potential interactions.
-Bioactivity data: Quantitative data on the activity of drugs against their targets, which can be used to assess the strength of potential interactions
National Institute of Health
-Dietary Supplement Label Database: Label and barcode pictures (& text) of different supplements available in the market
Data Science Approach
We used various classification algorithms to predict the risk of drug interactions. These include Random Forests to handle complex interactions between features and provide robust predictions, Gradient Boosting Machines (GBM) to improve predictive accuracy by combining multiple weak learners, Support Vector Machines (SVM) for high-dimensional classification tasks, and Neural Networks, specifically deep learning models, to capture intricate patterns in the data.
Evaluation
We used the mean F1 score, mean F2 score and mean area under curve to evaluate our reaction models. F1 and F2 scores are both harmonic mean averages of precision and recall that helped balance the trade-off between reaction and no reaction outcome. The main difference between the two scores is that the F1 score gives equal weight to both precision and recall, while the F2 score gives more weight to recall than precision so in the end we relied more on the F2 scores.
Key Learnings & Impact
A drug interaction is a change in the way a drug acts in the body when taken with certain other drugs, foods, or supplements or when taken while you have certain medical conditions. Other than unwanted side effects, and allergies, interactions could cause a drug to be more or less effective, cause side effects, or change the way one or both drugs work. Infact, the Food and Drug Administration’s Center for Drug Evaluation and Research estimated that 2 million serious adverse drug events occur in the USA each year, killing approximately 100,000 people
Acknowledgements
Course
Data Science 210. Capstone , Summer 2024More Information
