1 of 17

ADHD Screening Tool

Final Presentation

2 of 17

ADHD - an honest list of fears and confessions

TEDx

Cal State LA

Not Just LIVING

but THRIVING with ADHD

Angela Aguirre

3 of 17

support early diagnosis of the ADHD by evaluating the risk based on demographic & behavioral factors

4 of 17

National Survey of Children’s Health (2017-2020)

109K

Surveys Completed

12K

Diagnosed ADHD

Data sources

Children (0-17)

Cash Incentives

Random Addresses

5 of 17

Missing Value

One hot encoding

Data preparation

Feature Selection

Remove values

Test/Train Split

6 of 17

True Positive (TP)	False Positive (FP)
False Negative (FN) Key Goal : Reduce error Child has ADHD and Model predict Healthy	True Negative (TN)

Modeling approach

minimize FN & maximize recall

Predicted value

0

1

0

1

Label 0 - Child is healthy, 1 - ADHD

Actual value

7 of 17

Selected Model - Random Forest

Other models

Gradient Boosted (.9)
KNN
Logistic Regression
Naive Bayes
Random Forest (.8)
Random Forest with different layers

Metrics

Confusion Matrix
ROC Curve

Class Imbalance

Mean Predicted

Probability

PR Curve

Confusion Matrix

TP 14,920	FP 2818
FN 778	TN 1244

Mean Predicted probability

ROC Curve

PR Curve

8 of 17

Model - Training and Inference

RF Classifier

Calibrator

Training

Inference

Input

RF Model

Calibration

.72

.64

9 of 17

How can we share the results?

Data Prep Notebook

Model Train Notebook

App Notebook

Git

Jupyter on Heroku

Voila App Notebook

Publish our notebooks as dashboards for parents

dashboard

10 of 17

11 of 17

Thank you

12 of 17

Appendix

13 of 17

TEDx

Cal State LA

Not Just LIVING but THRIVING with ADHD

Angela Aguirre

14 of 17

Links

Application: Link
Git Repository: Link
Notebook: Link

15 of 17

Team Contribution

Prashant Dhingra - Build multiple modeling architecture - Gradient boosted decision Tree, Random Forest. Built matrix layer to evaluate model e.g. Confusion matrix, ROC, PR, Mean predicted probability. Build calibration curve to reduce over prediction.
Sebastian Urbina - Extracted the data set and performed the data wrangling. Developed the pre-processing scripts, to select the features, impute data, clean data and remove data sources. Wrote the documentation for the app. Provide feedback on modelling and UI
Jordan Thomas - Helped with initial EDA. Built baseline model pipeline. Built front end application notebooks. Built deployment process for front end application.
Joy First - conducted preliminary research (including project option identification, respective datasets, and reference materials), developed a product vision, led the delivery schedule and task assignment, coordinated team meetings and deliverables, conducted subject research and feature selection, and collaborated on data preparation, ML modeling and UX design.

16 of 17

References

“Economic burden of attention-deficit/hyperactivity disorder among adults in the United States: a societal perspective.” PubMed, 22 November 2021, https://pubmed.ncbi.nlm.nih.gov/34806909/. Accessed 28 August 2022
“Accurate Identification of ADHD among Adults Using Real-Time Activity Data.” NCBI, 26 June 2022, https://www.ncbi.nlm.nih.gov/pmc/articles/PMC9312518/. Accessed 28 August 2022.
“Predicting Children with ADHD Using Behavioral Activity: A Machine Learning Analysis.” MDPI, https://www.mdpi.com/2076-3417/12/5/2737/htm. Accessed 28 August 2022.

17 of 17

TP 17496	FP 242
FN 1663	TN 359

Gradient Boosted Decision Tree

TP 14,585	FP 3153
FN 736	TN 1286

Random Forest

RF

GB has .90 and RF has .82 accuracy.

But RF is better for FN

0

1

0

1

0

1

Predicted value