The NSCH is a household survey that produces national and state-level data on the physical and emotional health of children 0 - 17 years old in the United States.
In the NSCH survey data is a question that asks:
- Has a doctor or other health care provider EVER told you that this child has… Attention Deficit Disorder or Attention-Deficit/Hyperactivity Disorder, that is, ADD or ADHD?
In this project, I create a tool that can give the probability that a child from the NSCH survey has ADHD, but is
undiagnosed.
The final model has an accuracy of 0.90 -in other words, the model correctly classifies if a child is diagnosed with ADHD 90% of the time
The final model has a precision of 0.47 —in other words, when it predicts a child is diagnosed with ADHD, it is correct 47% of the time.
The final model has a recall of 0.90 —in other words, it correctly identifies 86% of all children diagnosed with ADHD.
The Next Steps I'd like to take are deploying the model to a place it can be interacted with, and adapting the model to predict other columns
Please see the project notebook for a full walk through of the data science process throughout the entire project.
ADHD is a developmental disorder associated with an ongoing pattern of inattention, hyperactivity, and/or impulsivity. The symptoms of ADHD can interfere significantly with an individual’s daily activities and relationships. ADHD symptoms begin in childhood and worsen in the teen years and adulthood
ADHD comes in many forms and severities, and teenagers and adults who are undiagnosed often don't even consider they may have ADHD if their symptoms don't perfectly match stereotypes of ADHD in the media.
As established above, Undiagnosed ADHD can be difficult for those with it. But why is it so common for ADHD to go undiagnosed?
This is a multi-pronged issue:
ADHD is not well understood. ADHD is still a subject of heavy research in the science/medical field.
The way ADHD is Diagnosed. To be diagnosed with ADHD, you would have to specifically bring it up with your doctor; this means that if you don't suspect you have it, you may never bring it up to your doctor on your own.
The way ADHD is portrayed People with ADHD are typically portrayed as young kids, who are extremely hyper-active and disobedient. However, hyperactivity is only one component of ADHD, and it expresses itself in various ways.
The NSCH is a household survey that produces national and state-level data on the physical and emotional health of children 0 - 17 years old in the United States.
In the NSCH data is a question that asks:
- Has a doctor or other health care provider EVER told you that this child has… Attention Deficit Disorder or Attention-Deficit/Hyperactivity Disorder, that is, ADD or ADHD?
For this project, I'm going create a tool that can give the probability that a child from the survey has ADHD, but is undiagnosed.
One distinction I'd like to make: This project is not about diagnosing kids with ADHD. Only a trained medical professional can diagnose someone with ADHD. This project and it's results are about spreading awareness.
The Data file can specifically be found here. Data in this project is the SAS data files from 2020, 2019, 2018.
To prepare the data for cleaning I:
- Removed columns related to ADHD
- Ensured all NaN values present in the dataset were addressed
- Performed a train-test split on the data and created a holdout set.
- Saved transformed/cleaned data as CSV files.
The full data cleaning notebook can be found here
I created 3 different types of models: Sklearn decision tree , Catboost, and a Keras neural network. My final model is a Catboost model.
For each model type, created a first simple model, and iterated from it, attempting to optimize the Recall and AUC scores.
For my final model, I attempted to address the class imbalance issue first, using the random oversampling technique, I then used the Optuna library to perform hyper parameter tuning on the model.
On the validation data:
The final model has an accuracy of 0.90 -in other words, the model correctly classifies if a child is diagnosed with ADHD 90% of the time
The final model has a precision of 0.50 —in other words, when it predicts a child is diagnosed with ADHD, it is correct 50% of the time.
The final model has a recall of 0.90 —in other words, it correctly identifies 90% of all children diagnosed with ADHD.
I've shown that it's possible to use machine learning to accurately identify kids with ADHD. Not only this, but I've shown it's possible to do this with just survey data, and if you can identify ADHD using survey data, it could be possible to identify other conditions, like depression, anxiety, and autism as well. If this were to be deployed, it could serve as a tool for those seeking answers, but unable to see a doctor to get an official diagnoses.
For a full narrative walk through and summary of all notebooks please see the ADHD Classification notebook, or see my presentation
├── Images <- Directory containing graphs and images from notebooks and presentation
│ └── ...
│
├── Model <- Directory containing the final model in cbm format
│ └── final_model.cbm
│
├── Notebooks <- Directory of detailed walkthroughs of project in Jupyter notebooks.
│ │
│ ├── Modeling <- Directory containing all Notebooks related to the modeling process.
│ │ └── ...
│ ├── Data_Cleaning.ipynb <- Detailed narrative walkthrough of data cleaning proccess.
│ │
│ └── EDA.ipynb <- Detailed narrative walkthrough of exploratory data analysis.
│
├── ADHD_Classification.ipynb <- Narrative summary of project in Jupyter notebook.
│
├── README.md <- Top-level README.
│
├── functions.py <- Helper functions for all Jupyter notebooks.
│
├── environment.yml <- The YML file for reproducing analysis environment.
│
└── presentation.pdf <- PDF version of project presentation.