DeepHeart

Expanding access to journaling for mental health.

We're in the midst of a mental health crisis:
‍Illnesses related to mental health are on the rise. Yet, we tend to ignore and underserve people suffering from depression, anxiety, and other "hidden" ailments.

Serving Therapists

In just 5 years, we'll be 250,000 mental health professionals short of projected demand. DeepHeart helps therapists reach more clients without being face-to-face.

Breaking the Taboo

Most people still feel uncomfortable talking about their mental health. DeepMind was created to make it frictionless for patients to journal everyday.

Applying Deep Learning

We give therapists a head start by applying deep learning techniques to classify emotions based on tone and words spoken. We recommend readings, too.

DeepHeart uses the latest machine learning techniques to help complement your practice:

App Architecture
Mobile Journal App

DeepHeart is available on both iPhone and Android. The front end is built in Reactive Native. Events are handled by Google Cloud Functions.

Our neural nets were created with Keras and trained on virtual machines set up in Google Cloud.

Emotion Classification

We detect a range of emotions — including Sad, Angry, Happy, Neutral, and Disgust — from speech recordings. We trained a Convolutional Neural Network (CNN) on a multilingual corpus of human speech. The corpus combines American English (RAVDESS), British English (SAVEE), German (Emo-DB), and Italian (Emovo).

Our model uses extracted Mel Frequency Cepstral Coefficients (MFCCs) from the speech recordings, which represents the envelope of the short time power spectrum. Each sound recording is broken down into 4 second slices. This allows us to predict emotions as they change over the course of journal entry.

We achieved a 65% test accuracy with a CNN model consisting of 2 1D convolutional layers combined with a max pooling layer and dropout for regularization. The model was implemented with Keras, Python, and Tensorflow.

Our model recommends articles by matching key words to topics on Psychology Today. Topic similarity is calculated using vectors from Glove's Wikipedia Word Embedding. We scrape content using Beautiful Soup in real-time.