Clinical records contain extensive data, but the links between a patient's problem, earlier decisions and later outcomes can be difficult to reconstruct.
A typical record can show every diagnosis, order, medication, measurement, and message. What it usually can't show, without hours of manual reconstruction, is how one problem changed over time: what evidence shifted its interpretation, what was tried, what the patient actually experienced, and whether things got better.
This is the third and final paper in our LIPI-KG series. Part 1 connected Lawrence Weed's problem-oriented record with Warner Slack's patient–computer dialogue. Part 2 defined the Problem Orientation as a computable unit. This paper closes the loop with a concrete proposal: use roughly 30 years of structured LEMR/iEMR data to train AI that reconstructs longitudinal problem solving.
The proposal is to build an inspectable representation of a problem's current state, supporting evidence and response to earlier decisions. It is not an autonomous diagnosis system.
What three decades of work actually built
The proposed LIPI-KG training approach draws on earlier problem-oriented informatics systems and their structured longitudinal records.
O-HEAP made Orientation explicit. Building on Weed, problem-focused knowledge navigation and the O-HEAP note put Orientation before History, Exam, Assessment, and Plan. Before asking what's new, the record asks where the problem stands relative to what was previously believed and planned. Repeated across visits, that produces a chain of states, evidence, decisions, and outcomes.
iEMR and LEMR preserved the structure. These systems kept discrete clinical elements, temporal instances, problem–plan relations, follow-up links, and longitudinal summaries. That matters enormously for training: a useful corpus has to record what belongs to which problem, when it was true, who said it, which decision it informed, and what followed. Text alone can't do that.
Terminology and reconciliation preserved intent. Later work in interface terminology, post-coordination, problem-list reconciliation, care-plan generation, and medication therapy management kept clinicians' and patients' original language while linking it to normalized concepts, and maintained the many-to-many relationships among problems, plans, and interventions.
Together, these systems provide a structured corpus of longitudinal clinical records.
Three things the AI must produce
1. A comprehensive, source-linked patient summary. "Comprehensive" doesn't mean long. It means organized around the patient's problems, and retaining what's needed to understand each one: recognition, evidence, relevant negatives, assessments, plans, interventions, patient experience, outcomes. Every substantive statement points back to its source and time. The summary is a readable synthesis for human review, not ground truth.
2. Candidate Problem Orientations. For each active, unresolved, recurrent, or historically important problem, the AI proposes an Orientation, using the form from Part 2: O = (P, t, S, E, A⁻). A good Orientation answers four questions compactly: Where does this problem stand now? What evidence supports that? Which prior plan is being evaluated? What uncertainty still needs review?
3. LIPI-KG. Each Orientation is an entry point into the Patient Orientation Knowledge Graph, a directed graph connecting the problem to the evidence, assessments, plans, interventions, patient experiences, and outcomes that explain its course. Provenance, time, confidence, and contradiction are all preserved. When sources disagree, both assertions stay until someone reviews them; nothing gets silently overwritten.
The core idea: graph tokens, not word tokens
Here is where this work departs most sharply from the usual approach of feeding chart text to a language model.
The proposed training unit is the problem-solving episode. An episode is a directed subgraph linking a problem state and prior plan to new evidence, assessment, action and a later Orientation or outcome. The aim is to learn these relationships, rather than treating each note or encounter as an isolated sample.
Navigation starts from the current Orientation. Instead of reading the chart from page one, the Orientation issues a typed query that walks backward across visits: the outcome being evaluated, the plan that preceded it, the assessment that motivated that plan, the evidence behind the assessment. That evidence chain is a meta-path, a reusable sequence of node and edge types. It separates relevant reasoning from the vast background record and leaves an auditable trail from every conclusion back to its source. Clinician-verified meta-paths can then become reusable learning assets in their own right.
Every clinical element becomes a metadata-defined token. A token here is the smallest retrievable unit of longitudinal meaning, defined by LEMR/iEMR metadata and its position in the graph. At minimum, each node token carries:
- original wording and normalized concept
- patient, problem-concept, and problem-instance identifiers
- event time, documentation time, and Orientation reference time
- source and provenance
- assertion status (present, absent, negated, uncertain, contradicted, missing, not applicable)
- O-HEAP role and problem lifecycle state
- confidence, relevance weight, and governance attributes
Edge tokens encode direction and meaning: supports, refutes, qualifies, precedes, revises, follows_plan, treats, evaluates, monitors, results_from, stabilizes, resolves, fails_to_resolve, and so on.
This distinction is fundamental. The same lab value or patient statement becomes a different token if it belongs to a different problem, occurred at a different time, came from a different source, or played a different role. Text embeddings can still enrich tokens, but they don't replace problem binding, temporal order, provenance, or directed relations.
Preparing the archive honestly
Normalization can't erase history. Evolving or duplicate problem labels get linked to stable concepts while keeping the original wording. Event time is kept separate from documentation time. Problem–plan links, revisions, follow-ups, and outcome adjudications are preserved.
Missing data needs special care. A field that was never used, a plan with no recorded outcome, or an intervention with no problem link must not be turned into a false "no." Missing, unknown, contradictory, and not-applicable are distinct states and stay distinct through training and evaluation.
A staged strategy
Each stage should advance only when the previous one is traceable and validated by humans, not just fluent:
| Stage | Output | How it's judged |
|---|---|---|
| 1. Tokenization | Typed node and edge tokens with problem binding, time, provenance, confidence | Token-class accuracy, relation direction, temporal accuracy, provenance retention |
| 2. Summary | Comprehensive, problem-organized, source-linked summary | Coverage, factual support, contradiction preservation, clinician review |
| 3. Orientation | Candidate O = (P, t, S, E, A⁻) | State accuracy, evidence sufficiency, prior-plan linkage, calibrated uncertainty |
| 4. Graph | Patient-specific LIPI-KG | Edge accuracy, temporal consistency, source traceability, completeness |
| 5. Learning | Outcome-associated favorable and unfavorable variance | External validation, clinical usefulness, bias and subgroup performance |
Two safeguards run throughout. Evaluation is split by patient and by time, not by random notes, so later information can't leak into earlier predictions. And the AI is allowed to abstain: when evidence is thin, contradictory, or temporally ambiguous, the right answer is an explicit "this needs review," with calibrated confidence and the supporting and challenging evidence on display.
From one patient to many
A single LIPI-KG explains one patient's problem-solving history. The bigger scientific value appears when many validated graphs can be compared. Graphs for the same problem can be aligned at meaningful landmarks (recognition, intervention, stabilization, recurrence, failure) and compared against outcomes.
The aim is to find longitudinal variance, not to measure compliance with a preferred pathway. Favorable variance might reveal sequences or patient adaptations associated with better outcomes. Unfavorable variance might reveal delayed recognition, missing follow-up, treatment burden, access barriers, conflicting plans, or repeated interventions that didn't help. These are associations, not causes; they're meant to generate hypotheses and decision support for prospective testing.
This is also where IPI and LIPI-KG complement each other. IPI identifies who is at risk across a population. LIPI-KG explains how an individual's problem evolved and supplies the structured paths that can then be compared across that population.
Why medication reconciliation still matters
Medications remain the primary demonstration across the series, even though they aren't the center of the graph. A medication is one intervention node. It becomes meaningful only when connected to the problem it was meant to address, the plan that introduced or changed it, what the patient actually obtained and took, what happened at home, and what followed.
Part 2's 714-entry example showed the gap: drug identity and start dates were nearly always recorded, while problem links, end dates, outcome adjudication, and reasons for stopping were sparse. That asymmetry is exactly why Slack's patient-at-home evidence is necessary. The record may hold the plan; often only the patient can explain actual use, benefit, side effects, cost, access, or abandonment. In LIPI-KG, prescribed, dispensed, available at home, reported, and actually taken stay distinct, and factors like cost, transport, insurance, or caregiver availability can enter as legitimate causes of a state change.
Humans stay in the loop
This is a reconstruction and orientation system, not a clinical authority. Clinicians and patients should be able to confirm, correct, reject, or reorient every summary statement, Orientation, and graph relation.
Patient-supplied evidence gets the same discipline as clinical evidence: source, time, consent, retention terms, and appropriate confidence. High-volume device data shouldn't dominate just because there's a lot of it, and clinician assertions shouldn't dominate just because they're professional. Weight should follow relevance and information value.
And the historical corpus carries its own biases. Thirty years of practice reflect the populations, workflows, and documentation habits of their time. The archive is valuable because it is unusually longitudinal, not because it represents every patient or setting. External validation, subgroup analysis, patient review, and prospective testing come before anything is used in care.
What this paper does and doesn't claim
This paper defines an architecture and a staged development plan. It does not report a trained model, a deployment, or validated causal findings. The full archive hasn't yet been normalized for this task, historical documentation is incomplete, and no home-generated data stream spans the whole corpus.
The proposed next steps are deliberately sequential. First, pick a bounded set of high-frequency problems, build a clinician-adjudicated reference set of summaries and Orientations, and test whether the AI can reproduce them with calibrated uncertainty and full traceability. Second, add patient-at-home medication, symptom, access, and device evidence for a prospective cohort. Only then compare accumulated graphs for outcome-associated variance.
Closing the series
Weed gave the record its organizing unit: the patient problem. Slack gave the patient a direct voice, including from home. O-HEAP, iEMR, LEMR, and the work that followed created the technical continuity between them.
LIPI-KG proposes a patient-specific graph linking problems, evidence, decisions and outcomes. The next step is to test whether AI can reconstruct those links accurately enough for clinicians and patients to review.
This post is adapted from "From Thirty Years of Problem-Oriented Records to LIPI-KG: Training AI to generate comprehensive patient summaries, Problem Orientations, and longitudinal knowledge graphs," Paper Three of the LIPI-KG series, presented at AMIA IPC, August 20, 2026.
Key references: Weed LL, N Engl J Med 1968;278:593–600 · Slack WV et al., JAMIA 2012;19:545–548 · 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) · Schrodt J et al., J Med Syst 2020;44:86 · Murali L et al., J Biomed Inform 2023;143:104403.