🇺🇸 Siddhu Pachipala

The approach

Risk assessment in mental health relies heavily on structured questions asked at a clinical appointment. Pachipala's project went at the problem from a different angle: applying machine learning to journal entries and analysing the semantics of the writing itself.

Language is where distress often shows first, and unlike a questionnaire it is produced without prompting. The Society for Science describes the method in one line and does not publish the datasets, the model, the accuracy or any clinical evaluation, so the performance of the approach cannot be assessed from this record.

Written language is also unusually rich as data. Word choice, tense, the use of absolutes, the shifting balance between first and third person and the topics a writer returns to have all been studied as correlates of psychological state, which is what gives a semantic analysis something to measure beyond simple keyword spotting.

Ninth place — and the Seaborg Award

Finishing ninth in the Science Talent Search carried $50,000 in 2023. Pachipala was also named the Seaborg Award winner, a separate recognition given within the finalist group.

The competition, founded in 1942 and in its 82nd year, took nearly 2,000 entries and named 40 finalists. The finalists shared $1.8 million; $3.1 million was distributed in total including school awards. The livestreamed gala was hosted by the broadcaster Soledad O'Brien.

The company at the top of the list

First place and $250,000 went to Neel Moudgal of Michigan for RNA structure prediction, second and $175,000 to Emily Ocasio of Virginia for measuring racial disparity in newspaper homicide coverage, and third and $150,000 to Ellen Xu of San Diego for a smartphone-based Kawasaki disease algorithm.

Between them and Pachipala sat work on q-calculus generalised into s-calculus, thyroid hormone treatment for traumatic brain injury, injectable microbubbles targeting stroke clots, machine-learning theory for gradually revealed data, and immune reactions to head trauma. Maya Ajmera, the Society's president, said she was "in awe of their creativity and conviction."

Why language models are used on this problem

Text-based risk detection is an active and contested research area. The argument for it is reach: far more people write — in journals, in messages, in forms — than attend a clinical assessment, and a signal that can be read from writing could identify people the system never sees.

The arguments against it are equally clear and are about consequences. False positives carry real cost in a mental-health context, consent and privacy are central rather than peripheral, and any tool of this kind is a prompt for human judgement rather than a replacement for it. Nothing in this record suggests the project was deployed or tested outside the competition.

The context explains why judges took the project seriously. Suicide prevention is a field with a persistent measurement problem: risk is assessed rarely, by people with limited time, using instruments that perform poorly at the individual level. Any approach that widens the window is worth testing, provided it is tested properly.

What is documented — and what is not

Firm: ninth place, the $50,000, the Seaborg Award, his age, his home town and the description of the method, from the Society for Science's March 2023 announcement.

Open: model details, accuracy, data provenance, ethical review, any publication, and what he has done since. This page reports the project as the competition described it and makes no clinical claim. Anyone in distress should contact local emergency services or a crisis line.