EP Mapping: ILAM research methods & limits

Local imports and one-to-many pairing

Import map folders containing point CSVs, EnSite DIF meshes and available waveform CSVs. Folder names suggest patient and ILAM/VT pairing, but the labels, pairing and VT-map relationship require review. One ILAM can have several VT references, and further patients/maps can be added without changing code. Different files are not automatically different VTs. Distinct morphologies are evaluated separately; complementary views/recordings may support one shared outcome.

Mapping files are processed in the browser. No supplied patient map is embedded in public assets or sent to a prediction API. Raw waveform Files stay in memory and are read lazily for a selected point; reopening a saved project may require reattaching those Files. Source hashes, point/mesh data and user annotations can be saved to the local browser and exported as a project. Exported projects and images may retain sensitive source information: keep them private and use deidentified inputs. Browser-local storage is not an authenticated clinical archive. The separate, optional ECG image review described below sends a consented ECG image to OpenAI; it does not upload mapping files or electrograms.

The importer preserves missing LAT as missing, native geometry and source validity/utilization flags. Display-only matrices are not applied to points alone. Supported original source annotation/label text and available coordinates are retained locally as reference-only records, distinct from the clinician’s new annotations; they are not discarded or sent to the prediction worker. Revealing them records reference exposure, and their presence raises a reference-marking review warning. Local project exports retain these records and may contain sensitive information. An ILAM filename never establishes the latest-component convention. Confirm chamber, surface, rhythm, pacing site, activation reference, annotation convention, window, units, orientation and platform. Unknown fields remain unknown.

Mapping PNG/JPEG/WebP images support viewing and clinician annotation, not automatic recovery of hidden surfaces, LAT, millimetres or patient-specific 3D anatomy. Review legends, timing windows and landmarks. Image annotations remain image-space observations; anatomical metrics require calibrated supplied geometry. Unsupported PDF/video or proprietary map formats need a compatible export rather than fabricated measurements.

Published rules and original computational choices

Irie uses the offset of local bipolar activity and eight equal time bins; crowding is >2 isochrones within a 1 cm radius. Aziz’s primary abstract specifies >3 within the same radius. The presets deliberately retain this difference. Conventional LAT annotations must not be silently treated as latest local electrogram completion.

What the engine can and cannot measure

Prediction requires confirmed latest-local-component timing in milliseconds, millimetre coordinates, a supplied triangular surface and adequate timed points attached within the configured tolerance. Mesh edges alone define connectivity. Invalid, degenerate or overlong triangles are excluded and gaps are not bridged. Nearby opposing walls, chambers and disconnected surfaces are never connected merely because Euclidean distance is short. Source surface identifiers additionally restrict neighborhood point collection when present.

Observed-bin counts approximate visually counted isochrones; they do not count contour boundary lines. Incomplete coverage may hide a zone. Nearest-vertex attachment and edge paths approximate a surface geodesic, with explicit projection and mesh-resolution limits. Region area is the sum of fully enclosed source triangles, not a projected circular area or inferred hidden surface. Local ms/mm values average offset-timing differences along available sampled mesh edges; they are substrate descriptors, not directly measured conduction velocity.

The 60/25/15 weights, connectivity search, neighborhood clustering and quality gates are original prototype decisions. They are not fitted to the demonstration patients or calibrated probabilities. Voltage is retained for inspection, not assigned an unverified scar threshold. Fractionation/late-potential data contribute only when supplied. No full conducting channel, line of functional block, exit, focal origin or ablation recommendation is automatically established. Latest activation alone receives no candidate status.

No usable candidates may mean insufficient sampling, an unsuitable window, incompatible annotation, unsupported geometry, a different activation wavefront or no region meeting this rule. It does not establish absence of VT substrate. The engine returns insufficient information instead of inventing missing observations.

Critical-isthmus candidates and exit/breakthrough

The current ILAM-only engine prioritizes candidate critical-isthmus regions from the existing crowding, later-connected timing and supplied electrogram features. The numerical weights and source geometry are unchanged. This target is a substrate hypothesis; each candidate retains an indeterminate circuit role and an uncalibrated rank score. No channel centerline, width, entrance, complete circuit or confirmed ablation target is reconstructed.

Exit/breakthrough is a separate assessment. It returns insufficient information and no assigned candidate IDs because the blinded ILAM input contains no VT-specific exit-localization method. One substrate can support several VTs and exits. An assumed seed in the generated phase illustration is not converted into a predicted exit, and an ECG regional hypothesis does not silently populate the ILAM assessment.

New prediction locks preserve both target assessments and their candidate identities, reasons and limits. Historical locks without target assessments remain role-unassigned; loading them does not retrospectively classify their regions as critical isthmuses or exits. Both targets are retained in the local validation index, without exposing withheld reference contents.

Reference comparison requires an exact target-role match: a critical-isthmus candidate is compared with explicitly annotated critical-isthmus references, while exit/breakthrough remains not predicted by this version. Entrance, channel, focal-origin and indeterminate reference labels do not substitute for either endpoint. Confirmed, probable and uncertain reference evidence remain distinguished. Missing target annotations are unevaluable, not negative findings. Target-specific summaries retain separate VT/patient grouping and do not turn spatial overlap into clinical accuracy.

The current engine is a written-rule research baseline, not a supervised model trained on uploaded pairs. Future development must preserve whole-patient splits, freeze both target definitions and model settings before held-out evaluation, and assess the two outcomes independently. Published ILAM/isthmus findings support the research rationale; they do not validate these original coefficients or establish performance for this application.

Exploration, locked predictions and reveal

Paired-case exploration is retrospective. In blinded workflow, the worker receives a copied allowlist containing only the selected ILAM’s point/mesh data, needed method metadata and point/mesh source digests. VT maps, reference annotations, descriptive source labels, registrations, ablation tags and arbitrary project state are omitted. The original supplied images must also be reviewed for burned-in reference information; unmasked reference markings make a case unsuitable for blinded testing.

Save the prediction, method settings, source/input digest, algorithm version and timestamp before revealing any reference from the patient. Revealing one VT exposes the patient for future assessments, including the other paired VTs; switching workflow or saving another prediction does not restore blinding. Only a genuinely unrevealed patient can support patient-level blinded demonstration. Predictions and comparison exports retain source/map identities and provenance.

Snapshot SHA-256 values support detection of accidental edits; a user controlling a browser can alter and recompute them. This is a local research workflow safeguard, not tamper-proof allocation concealment or examination security. Reference leakage through the supplied ILAM itself requires human inspection; software cannot certify that every embedded marking has been removed.

Reference evidence and separate VT comparisons

Annotate a reference region with role, confirmed/probable/uncertain certainty, mechanism and supporting activation, entrainment, termination, pace-map or other evidence. Confirmation belongs to the clinician and supporting recordings; the software does not promote an annotation to confirmed because it overlaps a prediction. Save each VT map and its evidence separately, allowing shared or different regions.

Every ILAM–VT pair needs its own reviewed registration. Identical supplied mesh coordinates and topology can support a proposed common frame, but names, point IDs or similar-looking views alone cannot. A rigid or uniform similarity fit uses at least three noncollinear corresponding landmarks, with VT as source and ILAM as target. The row-major 4×4 transform maps VT into the ILAM frame. A numerical fit remains unreviewed until the clinician checks anatomy, surface, units, landmark distribution and residual error; low landmark RMS does not certify accuracy elsewhere.

Anatomical comparisons require acceptable/good registration, matching patient and map identities, calibrated source geometry and an invertible rigid/positive uniform similarity transform. Shear, reflection, cross-patient transfer, image-only millimetres and unregistered overlays are rejected. The candidate is transformed into the VT source frame and compared over its supplied surface graph. Reported distance is an approximate surface separation from the candidate neighborhood to the reference region; area intersection/union is available only where source triangles support both regions. Similarity-scaled metrics refer to the registered target frame and need explicit physical interpretation.

Show every candidate–reference comparison, uncertainty and missing metrics. Patient summaries distinguish at least one covered reference from all documented VT outcomes. Complementary/same recordings are grouped as one VT outcome; an unresolved relationship has no asserted distinct-VT count. These are preliminary case demonstrations, not sensitivity, specificity, validated accuracy or treatment-success estimates. Any future learned method must split by whole patient, retaining all related maps, views and VTs together.

Map inspection and signal provenance

The comparison workspace keeps observed ILAM and a candidate-based generated VT phase illustration side by side. A third Recorded VT panel can be revealed after saving the current prediction in blinded mode; retrospective exploration explicitly records reference exposure. Mobile tabs provide the same three views without compressing their canvases. The selected recorded VT remains separately paired to its ILAM. ILAM and the generated illustration share their source bounds and camera. A recorded VT view is transformed into the same display frame only after acceptable reviewed same-frame, rigid or positive uniform similarity registration; camera motion is linked by default then. Source coordinates remain unchanged. Unreviewed maps stay visibly unaligned.

The mapping canvas projects supplied geometry and point timing, supports rotate/pan/zoom and click inspection, and keeps candidate overlays separate from original map annotations. Any display sampling affects drawing only, not the full analysis input. Reference overlays require appropriate registration; unregistered maps retain independent frames. No hidden chamber geometry is synthesized and no generic ventricular mesh is substituted. Imported vertices and source connectivity take precedence over ancillary screenshots. A point-only import stays a point cloud; without a usable source surface the generated map is withheld. Ambiguous competing surfaces or duplicate source identities are rejected instead of silently combined. Unverified display matrices do not establish anatomical axes. A recorded VT map is imported evidence; critical circuit annotations still require confirmation and supporting evidence, rather than being declared confirmed automatically.

Electrograms appear directly below the map. Select a mapped point, browse Previous/Next, or inspect the nearest actual source point within a candidate region; then select an exported channel and Show EGM. Candidate-point selection is a browsing aid, not proof of a critical electrogram or ablation target. Large signal files are streamed locally one recorded interval at a time and must be reattached after reload. Synthetic cavity examples have no recorded electrograms.

Source freeze-group/point identity, sample rate and confirmed channel alignment determine waveform timing. Reference-relative milliseconds and activation/reference markers are shown only when applicable; otherwise timing is from the start of the interval. Hover, touch or keyboard inspection reads the recorded sample without interpolation or signal interpretation. The display autoscales amplitude, which remains uncalibrated source units. Missing reference waves, placeholder exports, mismatched sample counts and unavailable intervals are flagged instead of drawn as invented flat or measured signals. Intracardiac electrograms are not relabeled as a simultaneous surface 12-lead ECG.

ILAM-only VT appearance hypothesis

An ILAM-only import can be reviewed and analyzed without any paired VT maps. After candidate ranking, VT Map Hypothesis constructs a separate generated phase field on the supplied chamber mesh. Choose a candidate and explicitly assume outward spread from it or reversed inward spread toward it. The candidate is an assumed breakout seed for this illustration; ILAM crowding does not establish that it is an actual exit, focal origin or complete reentrant circuit.

The original vt-appearance-hypothesis-0.1.0 experiment uses shortest paths along accepted supplied triangle edges, not a trained VT predictor. Equal graph distances receive equal normalized phase. No speed, tissue anisotropy, refractoriness, VT cycle length, activation milliseconds or ILAM-derived slowing is inferred. The 0–1 legend is dimensionless phase. Disconnected regions remain uncoloured, source geometry remains unchanged, and no hidden surface or true patient anatomy is reconstructed from points or images.

Generated maps are marked hypothetical, kept out of measured reference comparisons and the ILAM engine, and exported separately with candidate identity, settings, source geometry digest and assumptions. The comparison view automatically regenerates the selected candidate/direction illustration; obsolete source or candidate results are withheld. Its normalized phase is not numerically compared with recorded activation milliseconds. Generating this illustration does not reveal withheld VT references. Synthetic demonstrations use original schematic LV/LVOT and RV/RVOT cavity meshes with broad, smooth outlet transitions and separate chamber shapes. These are visibly labelled and are not supplied patient anatomy. Existing saved examples are not silently reshaped; load a fresh example to see the revised surface. Source-normal illumination smooths material lighting across existing triangles without moving source vertices, filling gaps or changing activation values. Lighting is independent of zoom, and very dense views use a cheaper material pass.

Export Validation Index produces a local patient-grouped research preparation file with current ILAM input digests, source digests, immutable prediction-lock versions/settings, whether those locks still match current input, and paired VT counts. Recorded reference IDs, registration and annotation counts are included only for patients whose references have already been revealed. The index contains no raw ECG/EGM samples or source point arrays. It is not a validated dataset or performance estimate. Preserve patient-level grouping in future training/validation splits; keep the separately exported generated illustrations distinct from observed recordings and verified outcome labels. Local checksums do not provide tamper-proof study blinding.

Preliminary ECG pattern review

A separate surface-ECG CSV import requires an explicitly calibrated time_s or time_ms column and simultaneous named leads. The current parser accepts 200–500,000 samples, 100–4000 Hz and uniform timing within 2.5%; missing samples and gaps are rejected rather than interpolated. ECG images and PDFs support local visual review and clinician-entered features. PNG/JPEG/WebP images additionally support the opt-in AI review below; PDFs remain local and require exporting a deidentified page as an image for that feature. No automatic tracing digitization, pixel-derived calibration or QRS detection is applied to images or PDFs. Automatic preliminary beat markers require calibrated waveform CSV. Lead placement, gain, pacing, noise and QRS boundaries require review.

The original amplitude-envelope beat detector uses a median baseline, a short smoothing window, a bounded refractory interval and explicit rate/quality gates. It estimates candidate beat markers, ventricular rate and RR variability. It is not the complete Pan–Tompkins filter/decision chain. Pattern labels use reviewable rate, QRS duration, regularity, P-wave and AV-dissociation findings; the complete Brugada wide-complex algorithm is not implemented. Wide-complex tachycardia remains unresolved when evidence cannot distinguish VT, aberrancy, pre-excitation, pacing or artifact.

The narrowly gated Betensky V2 transition-ratio comparison uses reviewed idiopathic outflow-tract VT/PVC morphology with V3 transition and a suitable sinus ECG, not scar-related VT. The ratio compares V2 R/(R+S) during VT with sinus rhythm; ≥0.6 favors an LVOT region in the source population. It supplies no anatomical coordinates, isthmus identification, complete circuit or proof of focal mechanism. Incompatible or missing morphology returns insufficient information; electrophysiologic mapping remains necessary.

Optional AI ECG image review

AI image review requires the separate signed AI-access session in addition to EP Mapping learning access. Unlock AI image review with the configured AI password; learning access alone cannot authorize a provider request. The server verifies both sessions before reading image content or consuming an analysis quota. The image-review consent checkbox starts unchecked. Confirm that both visible text and embedded metadata are deidentified before enabling it. While checked, the current supported ECG image and newly imported supported ECG images are automatically reviewed; Analyze or Retry can request another review. This is a session-only opt-in: uncheck it to stop further uploads. CSV and PDF files are never sent by this feature.

The application accepts PNG, JPEG or WebP images up to 8 MiB for AI review. It sends the image bytes and explicit consent through an authenticated, same-origin endpoint to the OpenAI Responses API. No application filename, patient assignment, mapping project, intracardiac electrogram, manual ECG features or prior review is added to that request. The image itself can still contain identifiers, which must be removed before import. Remote image URLs are not accepted. The server keeps its API credential private, applies request and response size limits, permits six requests per ten-minute quota window and times out provider processing after 60 seconds.

Review is an unvalidated educational image interpretation, not automatic waveform digitization or a calibrated measurement. It reports visible leads, image adequacy, preliminary rhythm, findings, alternatives and limitations. Unreadable or inadequate images have interpretation and localization withheld. Rate and QRS-duration estimates are withheld unless timing calibration is reported as visible; even then, verify them independently against the original ECG. Strict structured output and bounded values help reject malformed responses but do not establish diagnostic accuracy.

A sufficiently supported multi-lead VT image may receive a broad, explicitly uncertain region hypothesis with supporting features and alternatives. Otherwise localization remains indeterminate or not applicable. For scar-related reentry, the 2019 HRS consensus distinguishes the surface-ECG exit region from the diastolic corridor; ECG appearance does not locate a critical isthmus, exact ablation target, complete circuit or mesh coordinates. No AI finding is a treatment recommendation or a substitute for electrophysiologic mapping.

AI review remains separate from clinician-entered ECG features, the local CSV rules, ILAM candidate scores, generated phase illustrations and recorded-reference comparisons. It does not populate manual measurements or change a map prediction. Export clinician-entered review exports the manual review only. Changing the file or patient assignment, removing the assigned patient, revoking consent or leaving the page clears obsolete AI results and cancels pending client work; cancellation cannot recall bytes already transmitted.

The application does not persist ECG image bytes or AI results on its server; they remain in browser memory for the current view. The server retains quota-control records rather than image content. Provider requests use store:false and application responses use no-store, but this is not a promise of zero provider retention: OpenAI abuse-monitoring, image-input and applicable account data controls may still apply. Review the linked provider data controls before enabling uploads. Educational model only; not validated for clinical decisions or patient care.