Models4PT

Models4PT is an open research platform for building, curating, integrating, and maintaining computable population-level causal knowledge in physical therapy and rehabilitation science.

The project asks:

What do we collectively know?

Models4PT is intended to turn evidence from scientific literature and other admissible sources into structured knowledge that preserves meaning, provenance, uncertainty, disagreement, and human review.

Models4PT constructs and curates population-level scientific knowledge. It is not a clinical decision-support system and does not perform patient-specific diagnosis, prognosis, treatment recommendation, or Bayesian inference.

Scientific literature and research knowledge
                    ↓
                 Models4PT
       curated population causal knowledge
                    ↓
        Clinical Inference Engine (CIE)
 population knowledge + individual information
                    ↓
       patient-specific reasoning systems

  The Clinical Inference Engine (CIE) is a separate, existing repository whose
  patient-specific reasoning system remains under development. The distinction is
  defined in [System Boundaries](doc/project-foundation/SYSTEM%20BOUNDARIES.md).

  ## Development workspaces

  Use the repository's `Models4PT.code-workspace` for day-to-day Models4PT
  implementation. This keeps searches, agents, tests, and edits focused on the system
  that owns population-level knowledge.

  Use `~/Projects/clinical-inquiry-ecosystem.code-workspace` at milestone boundaries or
  when a change affects a cross-project representation, educational handoff, or
  software/data interface. Use `~/Projects/physiolog-simulations.code-workspace` only
  when work also requires physiological model development or simulation validation.

  Workspace membership does not create runtime or source-code coupling. Models4PT and
  CIE remain separate systems connected through explicit, versioned knowledge contracts.
  See [Integration Guide](doc/INTEGRATION.md) for the development rules and checkpoints
  that keep dedicated work usable by the rest of the ecosystem.

Models4PT is governed by several durable commitments:

• The canonical product is curated causal knowledge.
• Concepts, variables, measurements, mechanisms, evidence, and causal claims are distinct scientific objects.
• Semantic relationships are distinct from causal relationships.
• AI output is candidate knowledge; researchers remain the scientific authority.
• Accepted knowledge must remain connected to its evidence, assumptions, reviewers, and revision history.
• New evidence should refine knowledge without erasing disagreement or provenance.

See Foundational Principles for the full statement.

Models4PT is in an early research and software-design stage. The repository currently contains a tested Stage 1 domain experiment, not a deployable knowledge platform.

Implemented today:

• Python representations of concepts, variables, measurements, sources, evidence, proposed causal claims, and curation decisions
• a deliberately simple candidate-extraction experiment
• an explicit ontology-resolution boundary between extracted terms and proposed causal claims
• FastAPI health endpoints
• a minimal React health-check frontend
• backend tests and backend/frontend continuous integration

Not yet implemented:

• persistent storage or a shared ontology repository
• scientific CRUD or query APIs
• publication ingestion
• production AI extraction
• researcher curation screens
• knowledge integration, version history, or governance workflows
• authentication, authorization, or production deployment

The constraints in the current experiment—such as one concept per variable and binary proposed claims—are implementation hypotheses, not final scientific commitments.

src/models4pt/              Python package and FastAPI application
tests/                      Backend tests
frontend/                   Vite, React, and TypeScript scaffold
doc/project-foundation/     Governing vision and architecture documents
archive/                    Superseded historical planning material
.github/workflows/          Continuous integration

Python 3.11 is the development and container baseline. Node.js 20.19 is the frontend and CI baseline.

For environment setup, local commands, Docker usage, and validation, see Development Guide.

Quick validation after setup:

python -m pytest -q
npm --prefix frontend run typecheck
npm --prefix frontend run build

The next milestone is a narrow, provenance-preserving curation workflow:

source passage
    ↓
candidate concept and causal claim
    ↓
ontology resolution or explicit unresolved state
    ↓
researcher review with rationale
    ↓
persistent reviewed knowledge record

This vertical workflow will establish the repository and curation foundations before broader literature ingestion, visualization, collaboration, and downstream reasoning interfaces are attempted.

Current tasks, working definitions, and completion criteria are tracked in the Active Workplan. Design decisions awaiting principal investigator input are collected in the Curation Record Questions.

The longer research and software roadmap is described in the Software Project and Research Program.

Models4PT is the integrative hub between research literature and mechanistic knowledge from projects such as Physiolog, and downstream patient-specific reasoning in the Clinical Inference Engine. This position creates a need for a strong core of knowledge representations, data structures, and algorithms.

A possible LA4PT learning series was discussed as a way to build the project author's working knowledge of linear algebra while helping physical therapists understand the computational foundations of Models4PT. Linear algebra connects directly to graph representations, machine learning, embeddings, probabilistic models, dynamical systems, and simulation. However, matrix representations do not by themselves determine the scientific meaning of concepts and relationships or preserve evidence, provenance, uncertainty, review, and revision history. Models4PT therefore still requires explicit semantic and epistemic representations even when their implementations use linear algebra internally.

Current direction: retain LA4PT as a possible educational component of Models4PT rather than a separate project or domain. Its exact role—prerequisite, companion curriculum, design notebook, or orientation to the project's computational foundations—remains undecided. No LA4PT implementation or repository changes are planned yet; the idea is intentionally deferred for further reflection and evidence from continued Models4PT development.

Models4PT is released under the MIT License.

Sean M. Collins, PT, ScD — GitHub