GPTFX-the new era of machine learning in mental health care

Author(s): MSVPJ Sathvik

Originally published on Towards AI.

Mental health has remained one of the most challenging aspects of modern healthcare. A lot of people among us suffer from mental health problems and are looking for support and timely diagnosis without being troubled by being pointed out like it’s taboo.

The traditional methods of dealing with mental health care are confined to human interactions, inflexible diagnostic tools, and upcoming machine learning models that we cannot rely on due to lack of transparency.

But imagine if we had a promising AI system that could not only detect mental health issues with excellent precision levels but also give a structured analysis by explaining the reason behind the decisions it has taken, just like a paid therapist!

Transforming our dreams into raw reality, we have developed the revolutionary GPTFX. This powerful AI-based system leverages OpenAI’s GPT-3 technology, bettering how mental health disorders are detected, monitored, and explained.

Motivation: When machine learning models are trained on embeddings, it is very tough to get explanations if we use LIME and SHAP. Moreover, the explanations generated will not be continuous, and it is easy to extract words from the dimension numbers of the explanation result. So, we have devised a novel method to solve both.

The origin of GPTFX- A blending of transparency and detection

Some incredible explainable models like LIME(Local Interpretable Model-Agnostic Explanations) and SHAP(SHapley Additive exPlanations) offer diagnostic accuracy by producing results in the form of keywords or feature numbers, which are pretty helpful but sadly aren’t understandable for non-experts. This is one reason — they aren’t significant enough to be adopted in sensitive domains like mental health.

Here lands our savior, GPTFX, solving this issue by using the GPT-3 embeddings for both classification and generating the analysis with explanations and reasons. All you need to do is provide data in the form of text- social media posts.

This system now runs through machine learning and deep learning models trained on GPT -3’s embeddings, delivering results with an accurate prediction and clear text-based explanation of the results.

This powerful fusion of large language models (LLMs) and explainable AI (XAI) elevates GPTFX beyond existing frameworks, offering clarity to both clinicians and patients.

How does GPTFX work?

GPTFX has been designed based on the GPT -3’s embeddings. These embeddings make the data machine readable by converting the unstructured text data into numerical vectors. These embeddings are more advanced than the traditional ones since they capture the context of any word in the data, providing a good analysis of the psychological conditions in the text.

How is classification dealt with?

Machine learning models, such as Support Vector Machines (SVM), Logistic Regression, and Decision Trees, are used to classify mental health conditions like Thwarted Belongingness (TBe) and Perceived Burdensomeness (PBu). These models utilize GPT embeddings as a means of categorizing these conditions. The conditions indicate the presence of high emotional distress and are pretty essential to identify risks like suicidal self-destruction and social isolation.

Best model and rivalry

Our report analysis suggests that among all the GPT models that include Ada, Babbage, Curie, and Davinci, GPT-Davinci has been the best performer, achieving a whopping 87% accuracy in detecting Perceived Burdensomeness. We couldn’t be more proud that our model has exceeded popular mental health-focused models like PsychBERT and MentalBERT.

How does our model generate explanations?

We have fine-tuned GPT-3 to generate better explanations for each prediction, implementing it in our framework. For example, imagine a Reddit post that shows signs of Thwarted Belongingness; the model would generate a response like:

“This prediction was based on the phrases: ‘no one invited me anywhere’ and ‘I alienated myself.”

We are incredibly excited to share that GPTFX provides continuous, human-like explanations in natural language, bettering LIME and SHAP.

This would help build trust in the system by giving clinicians and patients a good understanding of how predictions are made.

How we monitor mental health using GPTFX

Mental health issues require regular monitoring since they often fluctuate; for this, we came up with an exciting idea to integrate our model GPTFX into AI-powered Internet of Medical Things (AI-IoMT) devices. To make it more user-friendly, we can embed GPTFX into wearable smart watches, mobile apps, and other IoT devices that analyze the user’s data, like speech, text messages, and social media posts, in real-time! This functionality helps in the early detection of mental health crises before they get any worse.

For example, assume a friend of yours is experiencing anxiety and texts you on chat about it; the system would act immediately by alerting the healthcare providers in real time. This helps monitor the user’s behavioral patterns regularly, offering a proactive approach by taking a significant leap in mental health care and intervening before the crisis escalates.

The best part is that GPTFX operates economically, relying on API calls to the OpenAI’s models! There’s no need for expensive infrastructure, making it easily scalable for adoption into healthcare apps.

Outperforming the other pioneering models

To bring out the best of the models, we have benchmarked GPTFX by rigorously testing it against leading models like BERT, ClinicBERT, and PsychBERT.

The results were outstanding, with GPTFX achieving superior accuracy(specifically GPT-Davinci) with 87.34% in classifying Perceived Burdensomeness.

GPTFX has also outperformed LIME and SHAP by generating highly readable and continuous explanations in contrast with how they(LIME and SHAP) struggled with contextual embeddings, giving keyword-based explanations that lacked coherence.

Additionally, GPTFX has scored a high 0.75, trumping ROUGE-1 scores. Furthermore, we’ve compared human-generated explanations against the ones generated by GPTFX, demonstrating a massive 80% similarity!

The results explain the importance of choosing the proper embeddings and models to achieve optimal analysis, diagnosis, and interpretation performance.

GPTFX’s horizon-The future of mental health AI

Providing accurate, explainable AI solutions in real-time, GPTFX paves the way for personalized patient-centered care and represents a radical transformation in mental health care. Now, mental health professionals and caretakers have better tools to identify risks using GPTFX, which combines credible diagnostics and cellular monitoring.

We see great potential in our framework that can further evolve by integrating it with deep learning models and connecting with sensitive areas like emotional sentiment analysis and cognitive behavioral therapy (CBT) platforms. The flexibility of the GPT embeddings makes it easier to adapt to new challenges, making it a long-term asset for the healthcare industry.

Conclusion

GPTFX is a superhero in the growing mental health crisis, combining powerful AI technology with the transparency and accuracy required for healthcare solutions.

It makes the world a better place to live by bridging the gap between human comprehension and machine learning, strengthening personalized diagnosis.

Join thousands of data leaders on the AI newsletter. Join over 80,000 subscribers and keep up to date with the latest developments in AI. From research to projects and ideas. If you are building an AI startup, an AI-related product, or a service, we invite you to consider becoming a sponsor.

Published via Towards AI