A problem list gives clinicians a current inventory of a patient's recorded conditions.
The list alone may not show when a problem was recognized, which evidence changed its interpretation, what treatment was tried or what followed. Those links are needed to reconstruct clinical reasoning.
In Part 1 of this series, we argued that medication reconciliation is where Lawrence Weed's problem-oriented record and Warner Slack's patient–computer dialogue meet. In this second paper, we turn that idea into a data model: a precise, computable definition of a Problem Orientation, and a description of how linked Orientations become the foundation of LIPI-KG.
First, three terms
Because these names are close, it helps to separate them up front.
IPI (Intelligent Phenotype Index) works at the population level. It identifies and stratifies people at risk, and its output is a cohort or risk category.
LIPI (Longitudinal Intelligent Patient-Centric Phenotype) is the evolving, patient-specific phenotype built from longitudinal clinical information.
LIPI-KG (Patient Orientation Knowledge Graph) is the directed graph form of LIPI. Entered through a Problem Orientation, it reconstructs how one patient's problem was recognized, evaluated, treated, stabilized, resolved, recurred, or failed to resolve.
Put simply: IPI asks who is at risk. LIPI-KG asks how this patient's problem evolved, what was done, and what happened.
The problem list as a projection
The key reframing in this paper is that the current problem list is not the structure itself. It is a projection of a deeper longitudinal structure, the view you get when you look at that structure from "right now." The richer structure underneath holds every historical state, every piece of evidence, and every plan that the current label summarizes.
That structure has to be organized around the patient problem, not around encounters, documents, medications, or isolated observations. Each problem becomes the center of a path connecting prior state, new evidence, assessment, intervention, patient experience, and outcome.
The Problem Orientation, formally
We define a Problem Orientation as a five-part object:
O = (P, t, S, E, A⁻)
- P is the normalized problem concept, while keeping the clinician's or patient's original wording.
- t is the reference time.
- S is the asserted state of the problem at time t.
- E is the evidence: supporting, conflicting, or missing.
- A⁻ is the prior assessment or plan that the current state is responding to.
That last element is what makes an Orientation more than a status label. It ties the present to a specific earlier decision, so every Orientation is implicitly asking: given what we planned last time, where do we stand now?
An Orientation can be produced during a clinical encounter or derived by AI from a comprehensive longitudinal summary. Either way, it becomes the entry point for retrieving the problem-centered graph.
Three design commitments
Separate identity from state. The problem concept should stay stable enough to support retrieval over years, while its state is always time-scoped. A problem can be suspected, recognized, under evaluation, treated, improving, stable, worsening, resolved, recurrent, or unresolved. The model doesn't force clinical judgment into a rigid state machine; it records each state as an assertion with time, source, confidence, and evidence.
Make relations explicit. The graph is built from a small set of element types joined by typed relations:
| Graph element | What it represents | Example relations |
|---|---|---|
| Problem concept | Stable clinical identity | has_orientation, related_to_problem |
| Problem Orientation | State at a reference time | precedes, revises, follows_plan |
| Evidence assertion | A finding, history, report, or observation | supports, refutes, qualifies |
| Assessment | Clinical interpretation | assesses, explains, changes_state |
| Plan or intervention | Action motivated by the problem | treats, evaluates, monitors |
| Outcome | Observed consequence | results_from, stabilizes, resolves, fails_to_resolve |
| Source | Clinician, patient, caregiver, device, system | asserted_by, observed_by |
Preserve provenance, uncertainty, and contradiction. A clinician's note, a patient's report, a caregiver's account, a lab result, a pharmacy fill, and a device reading are not interchangeable. Every assertion keeps its source and effective time. When two sources disagree, both are kept. A later adjudication may settle the question, but the evidence trail stays inspectable.
From a snapshot to a path
A single Orientation is a cross-section. Linked together, Orientations form a longitudinal path:
problem → prior Orientation → new evidence → current state → assessment → plan → subsequent evidence → next Orientation
Where the problem list shows one active label, the path can reveal the full sequence: a problem recognized and resolved, repeatedly stabilized and destabilized, refined into a more specific diagnosis, or never adequately addressed. And because it's a graph, several problems can share evidence or interventions without being collapsed into one.
Where the data comes from
Two pieces of prior work make this reconstructable. LEMR supplies the metadata: discrete elements, instances across encounters, temporal identifiers, problem–plan relationships, and continuity from visit to visit. O-HEAP supplies the recurring reasoning sequence: Orientation locates the problem relative to the prior plan, History and Exam add evidence, Assessment interprets it, and Plan records the next step. Repeated over time, O-HEAP naturally produces a directed chain of states, evidence, decisions, and outcomes.
The development goal is to train AI on roughly 30 years of LEMR/iEMR data so that an Orientation can retrieve its evidence and reconstruct the corresponding graph, with every assertion pointing back to its source and time. That training framework is the subject of Part 3.
A stress test: 714 medication entries
Medication reconciliation is not the model, but it is a revealing test of it. A medication is just one kind of intervention node, alongside labs, imaging, procedures, referrals, counseling, observation, or a deliberate decision to wait.
In one de-identified longitudinal record spanning 785 visits, we examined 714 medication entries:
| What we asked | Field | Populated |
|---|---|---|
| What action was recorded? | Drug identity | 714 (100.0%) |
| When did it begin? | Start date | 661 (92.6%) |
| Which problem motivated it? | Problem link | 428 (59.9%) |
| Was later access recorded? | Last refill | 347 (48.6%) |
| When did it end? | End date | 58 (8.1%) |
| What followed? | Outcome adjudication | 18 (2.5%) |
| Why was it stopped? | Discontinuation reason | 6 (0.8%) |
In this selected record, drug identity was complete, but links to the treated problem and later outcomes were less often populated. The example illustrates the problem–intervention–outcome links the proposed model is intended to preserve.
One caution: this is a single selected record, offered as an illustrative stress test rather than a population estimate.
Bringing the patient in
Slack's vision matters here too. Some decisive evidence exists only with the patient: what changed at home, whether the plan was feasible, whether symptoms improved, why something was altered or abandoned. In this model, patient-supplied assertions enter the relevant Problem Orientation with provenance and time, not as free-floating questionnaire answers and not merely as corrections to the clinician's list.
Limitations
The model is grounded in prior systems and a long-running archive, but it has not yet been validated across the full 30-year corpus. De-identification limits review of narrative evidence. The proposed states and relations need prospective validation, terminology governance, and review by clinicians and patients. And the model preserves provenance and uncertainty; it does not itself determine clinical truth or make treatment decisions.
The takeaway
A Problem Orientation records a problem's state, supporting evidence and relationship to an earlier plan. Linking Orientations over time provides a structure for reviewing the course of that problem.
In Part 3, we describe how to use three decades of problem-oriented records to train AI that builds these graphs, and what it would take to learn from them.
This post is adapted from "From Problem List to Problem Orientation: A longitudinal data model for constructing LIPI-KG," Paper Two of the LIPI-KG series, presented at AMIA IPC, August 20, 2026.
Key references: Weed LL, N Engl J Med 1968;278:593–600 and 652–657 · Meyers KC, Miller HJ, Naeymi-Rad F, Proc AMIA Symp 1998:325–329 · Naeymi-Rad F et al., US Patent 8,589,400 B2 (2013) · Naeymi-Rad F et al., US Patent 11,521,717 B2 (2022) · Tam VC et al., CMAJ 2005;173:510–515.