What is one specific practice or architectural choice that is essential for building predictive models in biological or healthcare settings, and why?

Most answers emphasize the importance of high-quality, unified, or sufficient data, while others highlight architectural simplification, scientific intuition, or experimental validation as essential practices.

10 answers

Answers

  • Andreas Cleve

    Co-founder & CEO · Corti

    Well‑labeled data

    Andreas Cleve says building deep models required a lot of data that is quite well labeled and well understood, which he frames as the founding ethos behind his company's approach.

    The European VC · Watch on YouTube
  • Rob Ferguson

    VP Technology and Strategy · Fireworks Ai

    Oracle Modeling

    Rob Ferguson says oracle modeling is essential, describing it as running every pure mutation of model combinations after a set of tasks to check what would be best, since the answer is always a mixture of models.

    Mind the Product · Watch on YouTube
  • Nworah Ayogu

    Partner · Thrive Capital

    Sufficient training data

    Nworah Ayogu says that before models can predict effects, there must be enough data to train them to make those predictions.

    Sourcery with Molly O'Shea · Watch on YouTube
  • Nirav Shah, MD, MPH

    Associate Chief Medical Informatics Officer, AI and Innovation · Endeavor Health

    Unified EHR with Semantic Layer

    Nirav Shah, MD, MPH says the fix for fragmented data is unifying the EHR into a single instance and streamlining the semantic data layer, because that unified, clean data foundation lets advanced tools like enterprise LLMs generate a cohesive patient narrative instead of forcing physicians to aggregate data themselves.

    The answer starts at 3:35.

    Becker’s Healthcare Podcast · Listen on Apple Podcasts
  • Pushmeet Kohli

    Chief Scientist, Google Cloud · Google

    Incorporate scientific intuition

    Pushmeet Kohli says to bake scientific intuition from biophysics and biochemistry into models, arguing it gives the model an unfair advantage and makes it more data-efficient because it doesn't have to relearn everything.

    Latent Space · Latent Space · Read
  • Richard Socher

    CEO · Recursive

    Parallel organoid experiments

    Richard Socher cites Peril Bio's parallel organoid experiments, saying the FDA approved them because how the organoids react to drugs and toxicity testing is more predictive of how drugs will interact in real human bodies.

    Peter H. Diamandis · Watch on YouTube
  • Colin LeClair

    Chief Executive Officer · Connections Health Solutions

    Single Data Lake

    Colin LeClair says leaders should get their house in order first, describing how his organization spent months building a single data lake of clean, discrete data across seven EHR locations so they could run reporting and analytics and feed AI agents.

    The answer starts at 21:48.

    Becker’s Healthcare Podcast · Listen on Apple Podcasts
  • Ramin Hasani

    Co-founder and CEO · Liquid Ai

    Simplify architecture

    Ramin Hasani says that the larger you make a model, the more you should simplify the architecture, because at scale less bias is needed inside the architecture.

    Latent Space · Watch on YouTube
  • Anuj Iravane

    AI Lead · Anterior

    Synthetic Data Pipeline

    Anuj Iravane recommends building a synthetic data pipeline, saying it lets you model edge cases and test workflows before going live in production without waiting on customer data.

    AI Engineer · Watch on YouTube
  • Eric Nguyen

    Co-Founder & CEO · Radical Numerics

    Tokenizing multimodal data

    Eric Nguyen says his company figured out how to tokenize modalities such as proteins, RNA, epigenomics and metabolomics so large language models can read them, which he says lets AI integrate data into a single system rather than modeling one siloed modality at a time.

    The Neuron: AI Explained · The Neuron: AI Explained · Listen at 17:44