You are on page 1of 27

D P L 3 0 2 M

FINAL PROJECT
R E P O RT
AI17BH2
Topics

Detecting emotion for


post-depression
treatment
Our Team

Nguyễn Đài Phương Trang Thạch Tiến Thịnh Lê Ngọc Khánh


SE173381 SE173658 SE172190

Nguyễn Hiếu Nhân Phạm Bùi Tiến Đạt Nguyễn Hùng Vinh
SE173393 SE173262 SE170557
Table of contents

BUSINESS CASE TECHNICALS

• Define problems • Data


• Status • Model AI
• Customers
• Values
• Features

TEAMWORK DEMO

• Plan
• Break-task
Business Cases

Why have us chosen to center


?
our attention on the subject of
depression?
Statistics

• According to WHO, more than


700,000 people die by suicide
each year, of which depression is
the cause of more than 50% of
suicide cases.

• Depression happens to everyone,


regardless of age or gender.
Users & Customers

Patient School

Doctor Company

Social media community Government


Project values

Online Post-Treatment Enhance Civic


Monitoring Communication
Patients and doctors connect The government can establish a
through an advanced web platform for individuals to
application, with AI support to interact and utilize AI
reinforce the doctor's rationale in technology to support, thereby
observing patients' mental states, enhancing the mental well-being
eliminating the necessity for direct and quality of life for the
interaction while ensuring citizens.
maximum convenience.
Features

Facilitates textual communication


1
and visual emotion tracking
between patients and psychiatrist.

Enables the creation of static and dynamic


2
dashboards, accessible through both the web
interface and downloadable files.

Support patient management at a


3
basic level.
Teamwork

Plans

• Divided into 3 modules operated


by 3 small groups.
• After every 1 week there will be
a sprint meeting to share
progress as well as knowledge
gained in the previous sprint
→ Discuss plans for the next
sprint.
Break-tasks
Break-tasks
Break-tasks
Break-tasks
Break-tasks
Technicals – Text model

Data
• Source: Hugging Face

• Overview: This dataset on emotional


expressions (e.g., "I feel the sky looks
beautiful today") is categorized into 6
emotion classes: happy, joy, fear, sad,
disgust, anger. Following the study in
(1), we reclassify these classes into 2
categories: positive or negative.

(1) Santos V, Paes F, Pereira V, Arias-Carrión O, Silva AC, Carta MG, Nardi AE, Machado S. The role of positive emotion and contributions of positive psychology in
depression treatment: systematic review. Clin Pract Epidemiol Ment Health. 2013 Nov 28;9:221-37. doi: 10.2174/1745017901309010221. PMID: 24358052; PMCID:
PMC3866689.
Pipeline

Data
Preprocess Data

Data
Solutions and Results
• TextCNN: Initial application scored 0.8191,
but misclassifications occurred due to word
order insensitivity, particularly in positive
sentences with negations.
• GRU: Subsequent application of the GRU
model resulted in a lower F1 score of
0.7591, attributed to the model's struggle
with longer sentences present in the
dataset.
• LSTM: Achieving 0.8689 F1 score, it
excelled on simpler sentences but faced
limitations with longer ones, exposing
constraints in real-world scenarios.
Solutions and Results

• Ensemble LSTM-GRU: Blending LSTM and


GRU for enhanced learning diversity,
achieving 0.822 F1 despite limited complex
sentences. Opted for real-world testing due to
time constraints, validating superior
performance on complex sentences. Finalized
as the ultimate model.
Technicals – Image model

Data
• Source: Kaggle (FER – 2013), self-
collected.
• Overview: This is a combination of
datasets:
• 1. Collected by class members using a
computer camera and sourced
• 2. From Kaggle.
Both are categorized into six main
classes: neutral, happy, fear, anger, disgust,
and sad. We reclassify these classes into 3
categories: neutral, positive and negative.
Data visualization

Data
Solutions
• CNN - VGG16: This method has several
main disadvantages
1. Limited data diversity in our collected
dataset.
2. Time-consuming processing leads to
delayed results, impacting user experience.
3. Diminishes user satisfaction by creating a
slower and less efficient interaction, reducing
overall user experience quality.

Using mediapipe bendshape to detect facial


emotion.
Workflows
Results
• Using mediapipe bendshape to detect facial emotion has achieved categorical accuracy at
91%
Results
• High F1 scores for positive and negative
classes, but lower precision in the neutral
Data class due to similarities with the negative
class. More diverse dataset needed for
improvement.
THANK
YOU FOR
LISTENING

You might also like