Data-Omnivorous Platform for Healthcare

03/19/2022 Compiled by Valmeek Kudesia; accumulated from real experience

Why a Platform?

See below for design for specific real experience and generalized design

Wish to Reach and Inform Every Clinical Decision

    • Numerous daily clinical decisions in varied settings, different member, patients, care-givers, & staff

    • Must address pain-points & needs that prevent analytics and prediction for day-to-day-use

      • Silos; teams on operating on similar but different datasets i.e. Conway's Law

      • Distinct dev/test/prod environments for datasets/data assets ⇨ additional overhead ⇨ lengthy iteration cycles, data validation & governance

      • Need faster iteration for analytics/data science/ML

    • Therefore, need capability that is flexible, extensible, and can "absorb" variety inherent to reality of clinical care

Imports Abilities of a "Data-Company"

Healthcare organization needs some but not all skills of data-savvy tech-company i.e must be "data-smart" not necessarily "database-smart". Therefore, must "import" following "data-smart" capabilities:

  • Quickly reach front-line end-users and addresses [[who-makes-decisions]]

  • Cost-effective ways to harness vast amounts of existing data

  • Quickly combine data domains to enhance value of data and give full view of business

  • Quickly improve data quality and expand utility of data

  • Provide “logical” layer to support data many diverse analysts vs only limited few

  • Rapidly learn and implement feedback from end-users (e.g. A/B testing, model dev)

  • Treat data as pluripotent asset; not “just a data warehouse/lake”

  • Separate storage from compute→supports many uses/service (not just fast reads and aggregations)

Critical Importance of Information Model

  • NOT building a warehouse, more like lake-house vs highly-organized lake

  • encodes subject matter expertise

  • integrated at patient/member and temporal granularity

  • Opinionated re "what is good" e.g. converge teams' actions, maximize work not done

Philosophy

  • [[data-ops]], CI/CD, open source

  • Data pipelines, transforms, models are “data applications”

  • No costly infrastructure/tooling

  • Incrementalism - accelerate with every new piece of work i.e. [[remove-causes-failure-better-job-with-less-effort]] and [[maximize-the-work-not-done]]

  • Deploying “new” is easy and controlled

    • Full architecture is replicated in “Dev place” and “Prod place“

    • Deploying to PROD is as simple as changing a config

    • Integrate w/ existing assets; engineer our own PRN

    • Data from all systems with speed, scale, reliability

    • Cohesive/contiguous data environment, warts included

    • High reliability and availability

  • Operate with the minimum amount of resources; encourages agility and efficiency; keeps things lean and pro-automation

Goals

  • Make all data (faults included) usable together to maximize scope of decisions inform-able

  • Accelerate use of data to generate insights/hypothesis to inform decisions at (or at nearly) continuous cycle time i.e. promote a learning organization and improve along things that never change

  • Reach every decision regardless of users’ skill to access and use insights

  • Accomplish the above with

    • High quality work

    • Minimal burden on data science and engineering

Objectives

  • Deploy a cloud-based advanced data platform that supports following capabilities via by modern rapid iteration and quick-to-value practices

  • Functional data pipelines & high quality data engineering

  • Descriptive analytics

  • Advanced analytics (e.g. ML, expert rules, rapid system learning)

  • Iterative data-to-information/concept modeling

  • Clinically-informed expert rules

  • Framework for A/IAI+human centaur system

  • Democratized self-service access as primary delivery vehicle for insights

Outcomes

  • Implements [[use-of-data-at-a-data-informed-org]]

  • Facilitates adoption of [[characteristics-of-data-informed-culture|characteristics of data informed culture]]

  • Aligns with [[identity-and-behaviors-clinical-organization | identity of clinicians]] thereby enables [[high-performing-clinical-model]]

  • Achieves [[vision-for-use-of-data-clinical-org]] and [[data-informed-compassion-guided-healthcare-org|data informed and compassion guided healthcare organization]]

Specific Design

Generalized Design