Niklas Bühler

Escaping local optima on a random walk through life.

PhD candidate at the Chair for Artificial Intelligence in Medicine, TUM, working on representation learning for medical imaging and multimodal disease risk prediction.

🩺 RisQ: Learning the Shared Structure of Human Health

A multimodal framework for disease risk prediction that learns a unified representation of a person's health from genetics, biomarkers, medications, lifestyle, and longitudinal clinical history, queryable in natural language for any disease over any time horizon. Trained and evaluated on ~500,000 UK Biobank participants, RisQ outperforms disease-specific models, multi-disease frameworks, and tabular foundation models, generalizes to unseen diseases and time horizons, and its learned representation recovers established disease genes and surfaces new candidates such as OMA1.

Preprint GitHub

🩻 Large-scale Radiograph Pre-training

Master's thesis on self-supervised MAE pre-training over 600,000+ clinical radiographs, cutting the need for labeled data across downstream medical imaging tasks.

Thesis Poster

🧬 Functional Gene Embeddings

Extracted functional gene embeddings via PCA, autoencoders, and a variational tensor factorization model, and evaluated them on GWAS trait prediction.

Report Slides

🧠 Connectome Informed Attention

Predicting tau spreading behavior in Alzheimer's patients from brain connectivity maps and tau PET scans.

Presentations 1 2 3 4 5 6 Poster GitHub

🧊 Bayesian Deep Learning – A Stochastic Dynamics Perspective

Overview of Bayesian deep learning through the lens of stochastic dynamics, covering the key methods used to train Bayesian neural networks.

Report Presentation

🧬 Interpretable Mechanistic Models for Predicting Tissue-specific RBP Expression

Predicting tissue-specific expression patterns of RNA-binding proteins with interpretable mechanistic models, mostly corroborated by literature and experimental data.

Sketch Report

🗣️ Crosslingual, Language-independent Phoneme Alignment

Bachelor's thesis applying cross-lingual, multilingual techniques to phoneme alignment, using hybrid HMM/ANN systems bootstrapped across languages.

Thesis Slides