Is Intel e-medix legitimate? Here is the whole system, on the record.

Intel e-Medix is a software-assisted clinical reporting and health-literacy system. It accepts a medical document or medical image, extracts and organises the clinical content, and produces two distinct outputs: a structured clinician-facing report and a plain-English patient explanation. It is an assistive reporting layer — it does not replace interpretation by a qualified clinician, and final clinical responsibility always remains with the treating physician.

The architecture is not a single model. It combines deterministic logic (file validation, category heuristics, storage, retention, audit), one or more vision/language models accessed through an external inference gateway with a fallback provider chain, a fixed output schema that constrains what a report may contain, and application-level checks and cross-checks. Different steps are performed by different components.

Accepted inputs are PDF documents, JPEG/PNG/WebP images and MP4/MOV video. The system does not accept DICOM (.dcm) files or complete multi-slice CT/MRI studies, and it does not read PACS. It therefore cannot independently inspect a raw imaging volume to detect, measure or stage a primary pathology. Where the input is a written report, the system analyses the reporting radiologist's or pathologist's own text. Where the input is a single photographic image, a vision model describes visible features — descriptive assistance, not a validated image-diagnostic function.

The modality is never assumed from the filename alone. A text/filename heuristic and a vision classifier are run and compared across twelve categories. Agreement raises the confidence label; disagreement lowers it and the interface asks the uploader to confirm the category before analysis proceeds.

Planned: explicit machine-readable anatomical region and laterality as separate validated fields; DICOM ingestion, viewing and metadata extraction.

Principle enforced in the reporting instructions: the system must not compensate for inadequate source quality by increasing diagnostic certainty.

Every clinician-facing report contains a mandatory Limitations section covering technical limitations, motion artefact, missing data and any factor affecting confidence, with an explicit “None — fully diagnostic” option. Where the source is unreadable or incomplete, the report states this and recommends repeat imaging or clinical reassessment rather than producing an interpretation.

Planned: a quantitative pre-analysis quality gate scoring resolution, focus, exposure, contrast, field of view, artefact, obstruction and anatomical coverage, with a hard block on interpretation below threshold. Today this assessment is narrative, not measured.

Findings are not free text. Every finding is emitted into a fixed schema with required fields: precise clinical term, measurement with unit and reference range where present, plain-English meaning, clinical significance, severity, an evidence quote taken verbatim from the source document, and a supporting guideline reference.

Reporting instructions require calibrated language and forbid asserting certainty the evidence does not support. Accepted certainty ladder: diagnostic of / highly suggestive of / consistent with / suggestive of / possible / indeterminate / cannot be determined from available data.

Planned: a machine-enforced OBSERVED / INFERRED / SUSPECTED / NOT ASSESSABLE flag on each finding. Today the distinction is expressed in wording, not as a validated structured field.

Severity is a separate graded field per finding and for the study overall, distinct from the impression. Exactly one triage level is assigned with a one-sentence justification, mirrored into a structured field (routine / soon / urgent) that drives dashboard prioritisation and emergency guidance. Recommendations separate additional testing, specialist referral, follow-up and clinical correlation.

Triage is derived from the clinical significance of the findings, not from a keyword match, and not every abnormal finding is escalated. Where emergency status depends on symptoms not present in the uploaded material, the report states that this cannot be assessed from the input.

ICD and procedure coding are not generated. No coding module exists.

Category detection selects a specialty-appropriate reporting template, so ophthalmic material is reported in ophthalmic terminology and radiology material in modality-specific radiological terms. The classifier is cross-checked precisely to prevent a generic template being applied under the wrong specialty label. A low-confidence classification is surfaced for human confirmation rather than silently accepted.

Planned: a hard modality-validation gate that refuses disease-specific interpretation logic until modality, region and laterality are confirmed.

Reports cite recognised frameworks (ACR Appropriateness Criteria, Fleischner, BI-RADS/LI-RADS/TI-RADS, ADA Standards of Care, AHA/ACC, NICE, KDIGO, WHO) to explain the classification framework used.

Known limitation, stated plainly: citations are produced by a language model and the system does not currently retrieve or verify the cited source. A citation must be treated as an indication of the framework applied, never as proof that the conclusion is correct. Reviewing clinicians should verify any citation they intend to rely on. Planned: a curated, retrieval-backed reference store so that only verifiable citations can be emitted.

Two outputs are produced from one analysis. The clinician report follows structured-reporting convention (patient/report information, examination and modality, indication, technique, source information, comparison, findings, impression, severity, recommendations, limitations, review status). A separate compact clinical PDF targets one to three pages, dense and prioritised. The patient explanation is kept separate and covers summary, findings in plain English, likely symptoms, recovery guidance, specialists, lifestyle guidance and references.

Every report carries the boundary statement that the system organises and explains an existing medical report, is not a diagnosis, and does not replace a physician.

The system is instructed never to appear more capable by being more certain, and never to manufacture missing clinical information. Where evidence is insufficient the report states so directly. Missing fields are rendered as “Not provided”, “Not assessable” or “Indeterminate” rather than filled by assumption.

Intel e-Medix publishes no accuracy figure. No sensitivity, specificity, PPV, NPV, F1, AUC or overall-accuracy value has been established for this system, because no clinical validation study has been completed. Any such number quoted about Intel e-Medix today would be unsupported.

When validation is performed, performance must be reported multidimensionally: sensitivity, specificity, predictive values, accuracy, F1 and AUC/ROC where the task is classification, calibration, false-positive and false-negative rates, subgroup performance, inter-reader agreement, report completeness, and diagnostic agreement with qualified specialists.

Aggregate accuracy is misleading on its own: a system can score highly overall while failing on rare but clinically critical conditions. Evaluation must report clinically significant failure modes separately from average performance. Reference standards and physician–system agreement studies are planned, not completed.

Monitored error classes, with false negatives and clinically dangerous errors given priority: missed critical findings, false positives, over- and under-diagnosis, incorrect severity, incorrect laterality, incorrect modality, incorrect coding, unsupported certainty and internally contradictory statements.

Today this is a manual process: outputs are sampled and reviewed against the source document, and physician and patient feedback is logged and folded into template revisions. Every report is traceable to the uploaded file and the uploading account through the audit trail. Planned: a structured error register with rate tracking per error class and per modality.

Planned — not yet implemented. The intended control is an automated pre-release scoring pass across accuracy, completeness, evidence support, clinical reasoning, severity, laterality, modality correctness, recommendation quality, triage correctness, uncertainty calibration, internal consistency, safety and conciseness, with a target of ≥9.5/10 on every applicable dimension before release.

Interpretation of the 9.5/10 target: it is an internal writing- and quality-control threshold. It is not a diagnostic-accuracy claim and must never be presented to users as “95% accurate”. A report that cannot reach the threshold on evidence must be flagged, not padded with invented content.

System analysis → structured report → quality and safety checks → physician review → final clinical report.

Physician-review notices are carried on report outputs and PDFs. Planned: a software-enforced review queue with named reviewer, sign-off state and locked post-signature versioning. Until that exists, review is an organisational process obligation, not a system control. System output does not constitute a physician diagnosis under any configuration.

Illustrative only; not a real patient case and not a diagnosis. Given a fundus photograph with features suggestive of diabetic retinopathy, correct system behaviour is: classify the modality as retinal/fundus imaging; describe the observable features; assess retinopathy severity and macular findings separately; state whether neovascularisation can or cannot be identified on the available image; decline to assert centre-involving macular oedema, because centre involvement requires OCT; and recommend OCT and ophthalmological assessment. Where image quality prevents assessment, the correct output is an explicit limitation, not a graded severity.

Intel e-Medix's contribution is consistency and communication, not autonomous diagnosis: a fixed structured schema instead of free text, evidence quoted from the source, modality-aware templates, explicit limitations and calibrated uncertainty, separated clinician and patient outputs, triage prioritisation, an audit trail linking every report to its uploader, and automatic 48-hour deletion of source files and analyses. No claim is made that the system outperforms clinicians or other systems; no comparative study has been run.

Intel e-Medix is designed to maximise the clinical quality, consistency and safety of software-assisted medical interpretation and report generation. Its actual diagnostic performance must be established through appropriately designed clinical validation against qualified reference standards. No single accuracy percentage should be used to represent performance across all diseases, modalities, populations and clinical scenarios.