Longitudinal Phenotypes and Disease Trajectories at Population Scale
summary
Population-wide Danish healthcare data is condensed into recurrent longitudinal phenotypes by testing directional co-occurrence and building linear trajectories from 7–10 million patients across a 40-year registry, then linking those trajectories to genomes and multi-omics. Methods include directionality exact tests, LSTM hourly prediction, deep-learning embeddings and an autoencoder; trajectories often predict ICU mortality better than 24‑hour vitals alone. Findings: many patients show mixed etiologies (≈ two thirds), replication fails across US data due to billing-driven coding density, and a public trajectory browser and prescription analysis of 1.1 billion prescriptions are produced.
AI-written summary · about
chapters (10)
- 0:00Start
- 1:38Disease trajectories are driven by different factors, including treatment and other exposures
- 2:22Combining deep phenotypic EHR data with registry data
- 7:37Disease trajectory network of epilepsy patients
- 7:55ICU patient mortality prediction from machine learning based aggregation of time scales
- 8:40ICU mortality prediction performance
- 12:33Summary: prediction performance
- 17:06DTB trajectories for Down syndrome
- 22:24Data set reconstructions
- 25:38A population-wide health data
related talks
01h 51mMorning Session: Genomics, Medicine, Epidemiology and Data Interpretation
39mSimulating a Million Patients: Realistic Health Data Generation
20mData-Driven Approach in Treating a Medical Condition
07mBig Data and Cardiac Rehabilitation
18mLongevity & Lifestyle
06mPassive Health Monitoring Using Large Scale Mobility Data
16mThe Secret Life of Data - Part 3
08mAddressing Big Data challenges for Real time Personalized Medicine