New AI Model Predicts Life Outcomes from Health and Career Data

Artificial intelligence continues advancing into uncharted territory, now claiming the ability to peek into people’s futures. Researchers from the Technical University of Denmark have developed a machine learning algorithm that can predict various life outcomes, like income and mortality risk, based solely on patterns gleaned from Danish national population registers.

A recent study in Nature Computational Science demonstrates AI’s expanding potential to extract insights from diverse datasets like work histories, salaries, and hospital visits. However, realizing any practical applications requires navigating complex questions around data privacy and algorithmic bias.

Training AI to Understand Life Stories

The researchers turned to a class of AI algorithms gaining widespread fame recently – large language models. The same methods powering chatbots like ChatGPT acquire language skills by “reading” massive volumes of text. This infuses them with understanding to then generate written content on demand.

The study’s team leader, Sune Lehmann, wondered if similar models could comprehend the sequences forming life stories. Just as word order matters in sentences, so does event order shape personal narratives. For example, a cancer diagnosis right after taking a job with rich health benefits will play out much differently than the reverse scenario.

Lehmann’s team synthesized a unique dataset from Danish national registers tracking every citizen’s employment status, salary, and hospital visits over decades to test this. Each discrete event, like changing jobs or a doctor’s appointment, became a line in an individual’s chronological “life story.”

This transformed population-level dynamics into a corpus for AI to uncover patterns. For instance, the model might learn those who never attend university tend to have lower lifetime incomes. The researchers named their model “life2vec” and trained it on the entire populace’s records from 2008 to 2016.

Predicting Mortality with Decent Accuracy

After teaching life2vec data story comprehension, the scientists tested its predictive capacities. In particular, they wanted to know if it could forecast a grim life event – whether someone had died prematurely by 2020 based on earlier records.

Remarkably, the AI’s predictions aligned with reality around 78% of the time. Beyond simply indicating higher risk for older folks, the model consistently flagged factors like mental illness, low socioeconomic status, and lack of social connections. It even registered the elevated likelihood of men dying early compared to women.

That said, life2vec struggled with unpredictable demises from accidents or sudden heart attacks. Still, the algorithm’s overall performance notably exceeded previous efforts at data-driven fortune-telling.

Intriguing Possibilities Alongside Challenging Questions

The Danish team found their AI could also anticipate other life attributes beyond mortality, like extroversion. However, associations between jobs like hairstylists and highly social personalities seem pretty obvious. The more integral question is whether life2vec genuinely uncovers deeper insights or just scratches the surface, connecting easily correlatable points.

Critically assessing what, if anything, this technology can usefully contribute requires confronting a slew of pressing issues. A top concern centers on whether patterns within a homogeneous Danish population would replicate across other societies or cultural contexts. It remains hypothetical if the model encapsulates any universal life story logic.

Additionally, biases encoded in the data that AI models train on can severely restrict reliability and lead to unintended harm if applied inappropriately. As just one example, the overdiagnosis of schizophrenia among Black individuals could falsely paint them at higher risk for premature death if an algorithm inherits skewed statistics. Any real-world application demands meticulous scrutiny to avoid unfairness or misleading predictions.

The Path Forward for AI-Assisted Life Analysis

While the study’s conclusions sound almost dystopian, author Sune Lehmann believes models like life2vec could eventually assist people in positive ways. In one case, identifying disease risk factors could motivate beneficial lifestyle changes. However, truly unlocking AI’s potential, even in narrow domains like Danish public health, necessitates confronting gnarly questions around ethical data use.

As Lehmann notes, “The best way I can think of to start this discussion is by forming an image of what’s even possible.” Their research undoubtedly expands the imagination of what AI can achieve by mimicking how people interpret the world. However, guiding further exploration requires charting solutions to emerging sociotechnical challenges alongside showcasing computational innovation.

If algorithms keep excelling at prognostication, at what point do data privacy rights and determination over our own futures come into question? And does their effectiveness convey any profoundly revealing messages about the human condition? As AI continues crossing new frontiers forecasting life outcomes, perhaps the most important revelations will be philosophical more than predictive.

Story Source: Reference Paper | Reference Article

Learn More: