Explainable machine learning to predict depression across phenotype definitions integrating genetic and environmental factors

Depression is a leading cause of disability worldwide, affecting millions of people, yet predicting who will develop it remains one of the most difficult challenges in modern medicine. The difficulty lies in the nature of the condition itself. Depression is not a single, uniform thing; it is a complex tapestry woven from many different threads. Some people experience it as a fleeting period of sadness, while others endure a lifelong struggle with severe symptoms. Furthermore, the causes are a tangled mix of biology and life experience. Scientists know that genes play a role, but they also know that sleep, financial stress, and social isolation are powerful forces. Because the definition of depression changes depending on who is asking and what questions they are asking, building a reliable computer model to predict it has been like trying to hit a moving target. Researchers have long wondered if the way they define the illness in their data changes the results of their predictions, and whether the genetic clues hidden in our DNA are useful only for certain types of depression or for the condition as a whole.

A team of researchers at the University of Edinburgh set out to solve this puzzle by testing how well machine learning models could predict depression when given different definitions of the disease. They turned to the UK Biobank, a massive database containing health information from over half a million volunteers, to examine nearly 153,000 people. Instead of relying on just one way to identify depression, the team created four distinct groups. One group represented people currently feeling depressed, identified by a standard questionnaire about symptoms in the last two weeks. The other three groups represented lifetime depression, defined by different levels of strictness: a broad screening based on core symptoms, a self-reported history of seeking professional help, and a strict combination of all these factors to ensure a high-confidence diagnosis. For every person in these groups, the researchers gathered a wide array of data, including their age, sex, sleep habits, financial situation, and a polygenic risk score, which is a single number summarizing a person's genetic likelihood of developing depression based on thousands of tiny variations in their DNA.

The researchers trained five different types of computer algorithms to learn from this data and predict who would be classified as having depression. They found that the most advanced algorithms, known as boosting models, performed the best, but the real story was in how the definitions of depression changed the results. When the models looked at lifetime depression, they were remarkably consistent. Whether the definition was broad or strict, the computer identified the same top five factors as the most important: being female, age, insomnia, self-rated health, and the genetic risk score. These models showed that genetic information added a small but meaningful boost to the prediction accuracy for lifetime depression. However, when the researchers switched the goal to predicting current symptoms over the last two weeks, the picture changed completely. The model for current symptoms relied much less on genetics and sex, and instead placed heavy weight on modifiable life factors, such as how often a person could confide in others and whether they were facing financial difficulties.

This shift revealed a crucial insight: the way we define depression dictates what the computer learns about it. The models trained on lifetime depression were highly transferable, meaning they learned the same rules regardless of whether the definition was broad or strict. In contrast, the model for current symptoms learned a different set of rules, focusing on immediate life circumstances rather than long-term vulnerability. The researchers also investigated whether genes and environment interacted, looking for cases where a person's genetic risk might be amplified by their life situation. They found strong evidence that the genetic risk for depression was more pronounced in people who suffered from insomnia, suggesting that sleep disturbance might act as a trigger that makes genetic liability more potent. They also found similar interactions with age and overall health.

The study concludes that there is no single "best" way to predict depression because the illness itself is not a single thing. A model designed to identify people at risk of developing depression over a lifetime will look very different from one designed to spot someone in the midst of a current episode. The findings suggest that while genetic data is useful for understanding long-term risk, it is less helpful for predicting immediate symptoms, which are driven more by current life stressors and social connections. By showing that the definition of the disease shapes the prediction, the research highlights the need for scientists and doctors to be precise about what they are trying to predict. If the goal is prevention, models should focus on lifetime risk factors. If the goal is immediate intervention, the focus must shift to current environmental and social conditions. This work does not offer a magic bullet to cure depression, but it provides a clearer map of the terrain, showing that to understand the future of mental health, we must first agree on what we are looking at.