Excited about our new paper! We developed a DL model to predict 1y mortality in the full Finnish population using A LOT of longitudinal information. This “digital aging marker” is very accurate, but with some interesting and important caveats www.medrxiv.org/content/10.1... - ThreadSky

andganna.bsky.social • 526 days ago

Excited about our new paper!

We developed a DL model to predict 1y mortality in the full Finnish population using A LOT of longitudinal information. This “digital aging marker” is very accurate, but with some interesting and important caveats

https://www.medrxiv.org/content/10.1101/2023.09.18.23295726v1

Comments

andganna.bsky.social•526 days ago

First, Kudos to Andrius Vabalas for leading this project! He never gave up even with such an overwhelming amount of data!

andganna.bsky.social•526 days ago

Our study utilized an unprecedented number of longitudinal features (>8,000) and data modalities with comprehensive, high-quality data collected in national registers for an entire Finnish population (N = 5.4 million) spanning up to 50 years.

andganna.bsky.social•526 days ago

We explored age and sex distribution in our data as crucial factors influencing mortality, with the majority of individuals dying at an advanced age The amount of longitudinal data available per individual over time gradually increased, with some registers starting more recently

andganna.bsky.social•526 days ago

The recurrent neural network (RNN) model included 2.9 million trainable parameters and achieved an AUC of 0.944 for 1-year mortality prediction, surpassing the baseline model trained solely on age and sex features, with an AUC of 0.897.

andganna.bsky.social•526 days ago

The RNN model showed good predictive performance across different causes of death, even when not seen in training data, i.e. COVID-19. There was a significant improvement over the baseline in predicting deaths from accidents or violence possibly due to socioeconomic predictors.

andganna.bsky.social•526 days ago

After removing the biological age effect, our model still has good predictive performance suggesting the potential as a digital marker of aging Model's performance decreased in older individuals: Heterogeneity increases as people age. Similar effect is seen for epigenetic clocks

Comments

Posting Rules

Reply