In a recent development in artificial intelligence (AI), scientists from Bilkent University in Ankara, in collaboration with psychiatrists, have pioneered an AI-based technology capable of detecting signs of depression and performing personality analysis using data such as voice, speech content, facial expressions and body language.
Hamdi Dibeklioğlu, assistant professor at the Department of Computer Engineering in Bilkent University, emphasized his lifelong focus on artificial intelligence research, particularly in the domain of automatic analysis of human behavior.
Termed as affective computing, the project analyzes human behavior utilizing machine learning techniques, discerning patterns from speech content, voice volume, tone, facial expressions and posture, according to Dibeklioğlu.
Highlighting the increased interest in affective informatics with the widespread use of AI models like ChatGPT, Dibeklioğlu revealed the team's recent strides in developing algorithms to gauge the severity of depression using AI.
"Typically, clinical psychologists and psychiatrists diagnose based on observations. Similarly, with AI, we aim to determine depression levels by analyzing various data – from facial expressions, tone of voice and speech patterns, to body language. While an expert conducts the interview, AI simultaneously processes and shares the data with the expert," explained Dibeklioğlu.
Dibeklioğlu further detailed that his research, initiated during his tenure in the Netherlands and continued upon his return to Türkiye, adheres to stringent ethical standards, securing approvals from both patients and hospitals at each stage. He emphasized the importance of data privacy and consent, reassuring that the system operates only with the explicit permission of individuals, safeguarding confidential and sensitive data through rigorous protocols.
"We term this as confidential or sensitive data, and its usage involves very meticulous procedures," told Dibeklioğlu, underscoring the seriousness and ethical considerations inherent in their pioneering work.
Use of clinical data
"We are trying to unravel the relationship between behavior and depression level. Our findings overlap with the theory. For example, there are different thoughts, 'my child cries a lot, is he depressed?' etc. However, in depression, the general expectation is dullness. In other words, the person does not want to communicate with anyone, he/she breaks away from the social environment. The patterns we capture are also like that. So, when you examine the literature, you come across behaviors of avoiding social interaction in depression. In other words, the model makes its own diagnosis. In this way, the AI can catch something that is overlooked," he explained.
Lie detection
Dibeklioğlu discussed a separate project centered around determining the extent of deception through various data points like sentences, vocal tone and visual cues. Ethical approval was obtained for these studies.
In analyzing diverse video materials, their focus was on gauging the authenticity of conversations, distinguishing truths from lies and cross-referencing these findings with multiple sources. The content of the speech underwent evaluation through "natural language processing" models, while the tone of voice was assessed via "frequency analysis."
Dibeklioğlu clarified: "Nevertheless, while artificial intelligence helped resolve this challenge, achieving a 100% precise prediction is unfeasible. However, we've achieved notably high success rates."
"It's crucial to comprehend that this lie-detection system cannot be directly employed in legal proceedings or decisions impacting someone's life due to inherent error rates. However, its applications extend to various domains. AI-powered lie detection might find use in settings like student or employment interviews. Unlike studies scrutinizing entire conversations for truthfulness, our approach targets understanding the levels of deception. It's a nuanced distinction; not every aspect of a conversation is deceitful, but not everything is an absolute truth either."
Personality detection
Dibeklioğlu elaborated on their assessment of personality across multiple dimensions, such as openness to the external world and innovation. They gather personality data through visual and auditory elements, interpreting them while engaging with individuals and aiming to impart this ability to machines. Despite the machine's enhanced capabilities for detailed observations and complex operations, the crucial factor lies in accurately training the algorithm.
He underscored the need for caution in the domain of human behavior, stating: "Absolute precision in this field, directly influencing daily life, may lead to significant issues if individuals are held accountable for mistakes. Ethical approvals in behavior analysis require meticulous examination. Our objective is for AI to assist us; however, this doesn't entail relinquishing all decision-making to AI while distancing ourselves from accountability."
Pain level detection
Dibeklioğlu highlighted their utilization of a comparable system to identify the "pain level," crucial in determining medication dosages.
He emphasized the potential significance of this detection, particularly in treatments for children and infants, stating: "With children and babies, it's often challenging to directly ask about their pain level. In these instances, we rely on interpreting facial expressions and behavior to gauge the degree of pain they might be experiencing."
