Conception & Design
-
External ventricular drains are among the most common neurosurgical procedures, with roughly 25,000 placed annually in the United States, yet day-to-day floor and step-down management of the indwelling drain is governed less by a single authoritative protocol than by fragmented society guidance, institutional habit, and the bedside judgment of whoever is covering.
The recognition burden falls disproportionately on junior residents, advanced-practice providers, and nursing staff who must integrate drain output, CSF character, intracranial pressure trends, the neurologic exam, and the underlying pathology into a real-time decision about whether a finding is benign, requires intervention, or demands immediate escalation, often at 3 a.m. with the attending off-site.
The failure modes are concrete and high-consequence: an occluded catheter mistaken for low intracranial pressure, overdrainage producing subdural collections or upward herniation, a tract hemorrhage missed in a coagulopathic patient, or a ventriculostomy-associated infection caught late.
AgentEVD targets exactly this recognition-and-escalation gap, structuring the same data a competent clinician would gather into a deterministic, guideline-anchored assessment so that the floor-level decision is standardized, auditable, and consistent regardless of who is at the bedside or what hour it is.
-
The contemporary evidence base for EVD care is anchored by the Neurocritical Care Society's consensus statement on the insertion and management of external ventricular drains, which synthesizes the available literature on placement, drainage, infection prevention, and weaning while explicitly noting that many of its recommendations rest on low-quality, single-center data (Fried et al., 2016). The conditions that most often require an EVD are governed in parallel by disease-specific guidelines that dictate the physiologic targets a bedside provider must hold simultaneously: the AHA/ASA aneurysmal subarachnoid hemorrhage guideline specifies enteral nimodipine, the post-bleed day 4 to 14 vasospasm window, pre-securing blood-pressure limits, and sodium management (Hoh et al., 2023); the AHA/ASA spontaneous intracerebral hemorrhage guideline defines acute systolic targets and anticoagulation reversal (Greenberg et al., 2022); and the Brain Trauma Foundation severe traumatic brain injury guidelines establish the 22 mmHg intracranial-pressure treatment threshold and the 60 to 70 mmHg cerebral perfusion pressure target (Carney et al., 2017). Infectious risk, the most consequential routine EVD complication, is addressed separately again by the IDSA healthcare-associated ventriculitis guideline, which standardizes the CSF sampling and empiric-treatment pathway (Tunkel et al., 2017). The knowledge required for safe EVD management therefore exists and is high-quality in places, but it is distributed across at least four independent documents written for different diseases and audiences, none of which integrates them into a single bedside workflow for the indwelling drain.
Beyond these targets, the literature documents both genuine management equipoise and quantifiable complication risk that a structured tool can encode directly. The optimal weaning strategy remains unsettled: a meta-analysis of rapid versus gradual weaning found no difference in shunt dependence or EVD-associated infection but a significantly shorter hospital and ICU length of stay with rapid weaning (Ebel et al., 2023), and a separate systematic review likewise found no clear weaning advantage while reporting that intermittent rather than continuous drainage was associated with fewer infections and catheter blockages (Palasz et al., 2021). Complication risk is increasingly stratifiable rather than merely descriptive: a single-center analysis identified platelet count below 120 ×10³/µL, white-cell count below 11 ×10³/µL, and intraparenchymal hemorrhage as independent predictors of catheter tract hemorrhage (Meyrat et al., 2024), and a review of cardiac emergencies in neurosurgical patients catalogs the Cushing reflex and related mechanisms by which raised intracranial pressure and neurosurgical stimulation produce bradycardia and hemodynamic instability (Chowdhury et al., 2015). Across these sources the recurring pattern is consistent: validated thresholds, risk factors, and physiologic targets exist in the literature but are left to the individual clinician to recall and integrate at the bedside, which is precisely the synthesis burden AgentEVD is designed to structure and standardize.
-
The project was deliberately built as an agent rather than a chatbot, and that distinction drove every architectural choice. A conversational model retrieves and answers; it is probabilistic, non-reproducible, and unauditable, which is disqualifying for a tool that touches escalation decisions.
AgentEVD instead holds structured state across a defined intake, evaluates that state against an explicit rule set, branches on the data, fires multiple concurrent concerns, ranks them by severity, and emits a fixed-format artifact (assessment, impression, prioritized recommendations, evidence-to-action ledger, SBAR, and SOAP note).
The decision path contains no language model at all: every recommendation is traceable to a specific encoded rule and a specific source, so the same inputs always yield the same output and any clinician can inspect why a given concern fired.
This was chosen over a retrieval-augmented chatbot because the publishable, defensible paradigm in clinical decision support is build-then-evaluate against guideline concordance and expert review, which requires a system whose behavior is fully specified in advance rather than generated on the fly.
-
The tool is architected so that no protected health information ever leaves the user's browser: AgentEVD is a single static HTML file with all logic running client-side, holding state in memory only, with no backend, no database, no network transmission, no cookies, and no analytics, so closing the tab destroys the session entirely.
-
Multiple layers of guardrail are encoded directly into the engine.
A hard gate prevents any recommendation from being issued until the minimum safe dataset (primary pathology and current drain status) is present, returning an explicit "insufficient data" state rather than guessing.
A set of caveat triggers surfaces automatically whenever the inputs would make the output less trustworthy: a missing or unassessable neurologic exam, the presence of sedation or other CNS-active agents that confound that exam, and corticosteroid use that can mask the white-count signal of infection, each of which prints a warning and, in the case of an unknown exam, caps the optimism of the rendered severity tier.
Red-flag physiology short-circuits to immediate escalation: a new fixed or dilated pupil, deterioration with new anisocoria, and analogous findings drive a herniation pathway to the top of the output regardless of what else is present.
Every numeric threshold is provenance-tagged in the code as either guideline-derived or institutional convention, centralized in a single tunable block so values can be audited and localized rather than buried, and every recommendation the tool prints carries its source and evidence tier so that no action is offered without an attached justification.
The system never claims diagnostic certainty; it frames outputs as considerations and consistently routes ambiguous or severe states toward human escalation.
-
Scope was bounded tightly and intentionally to keep the project rigorous rather than sprawling: adult patients only, the indwelling floor and step-down management phase of an already-placed EVD rather than insertion technique or operative decision-making, and a curated set of high-prevalence concern categories (drainage failure and overdrainage, intracranial hypertension, hemorrhagic and infectious complications, neurologic deterioration, the aneurysmal SAH vasospasm window, sodium and acid-base derangements, blood-pressure targets, bradycardia/Cushing physiology, and weaning) rather than an attempt to encode all of neurocritical care.
Feasibility was secured by the zero-dependency, client-side architecture, which removes server cost, deployment complexity, and data-governance overhead, allowing the entire tool to be hosted as a single file on free static infrastructure and embedded in an existing site.
The knowledge base was deliberately constrained to a fixed, verified corpus rather than live literature retrieval, which makes the behavior reproducible and the validation tractable within the timeline.
The remaining effort, the part that determines real-world value, was consciously allocated toward testing, source documentation, and evaluation rather than feature accretion, consistent with the project's emphasis on a defensible build-then-evaluate path.
Knowledge Framework
-
AgentEVD does not query live databases at run time; its knowledge is derived from a curated corpus of society guidelines, systematic reviews, and primary studies that are read, tiered, and aggregated by discrete decision rules.
Knowledge extraction is therefore deterministic and human-mediated rather than statistical: each clinically actionable statement (a numeric threshold, an indication, a recommended action, a risk association) was located in its source, assigned an evidence tier, and translated into an explicit conditional in the rule set, with the source citation and tier attached to the output it generates.
Tiering follows a fixed hierarchy in which an EVD-specific society guideline outranks a disease-specific guideline, which in turn outranks a review, systematic review/meta-analysis, or primary cohort.
Every threshold carries a provenance tag distinguishing guideline-derived values from institutional conventions, and conventions are flagged in the interface as tunable, so the boundary between what the literature dictates and what local practice sets is always explicit.
The result is that the tool's "knowledge" is fully inspectable, each recommendation maps to a named document and tier, and updating the knowledge base means editing a rule and its citation rather than retraining anything.
-
The linked folder contains the verified guideline-and-evidence corpus that the rule set is built from, spanning the EVD-specific and disease-specific society guidance that anchors the high-confidence rules and the reviews and primary studies that support the remainder.
At the top tier are the documents written specifically for ventriculostomy care, the Neurocritical Care Society's EVD insertion-and-management consensus and the relevant nursing care guidance, which drive the drain-handling, patency, leveling, and infection-prevention logic.
Disease-specific guidelines supply the pathology-conditioned rules: the AHA/ASA aneurysmal subarachnoid hemorrhage and intracerebral hemorrhage guidelines for blood-pressure targets, nimodipine and the vasospasm window, sodium management, and hemorrhage reversal and surgical indications; the Brain Trauma Foundation severe TBI guidelines for the intracranial-pressure treatment threshold, cerebral perfusion pressure target, and ventilation; and the IDSA healthcare-associated ventriculitis guideline for the infection workup. Supporting evidence includes systematic reviews on EVD weaning and on EVD use in SAH, a primary study on predictors of tract hemorrhage that informs the coagulation and platelet risk logic, and a review of cardiac emergencies in neurosurgical patients that grounds the bradycardia/Cushing pathway.
Each document in the folder corresponds to one or more named rules in the engine, and the in-tool evidence-to-action ledger names the specific source behind every recommendation it issues.
-
AgentEVD is implemented as a single self-contained HTML document with inline CSS and vanilla JavaScript and no external runtime dependencies beyond two web fonts, chosen so the entire application is one auditable artifact that runs anywhere a browser does and can be hosted as a static file. Application state lives in a single in-memory object that the intake controls mutate directly; the interface is built programmatically from small helper factories (button groups, sliders, conditional reveals, fishbone lab layouts) so that the form, its conditional logic, and its abnormal-value styling are generated consistently rather than hand-written per field.
The clinical core is a rules array in which each rule is a self-describing object carrying a firing predicate over the state, a severity function, differential considerations, immediate and subsequent management lists, a plain-language rationale, a one-line synthesized summary, a short impression fragment, and its source citation; rule fields may be static or functions of state, which lets recommendations adapt to inputs (for example, blood-pressure targets that resolve by pathology, or hemorrhage management that changes with coagulation status).
The execution flow is deterministic: validate the hard gate, assemble caveats, evaluate every rule predicate, sort fired rules by severity, short-circuit on red-flag physiology, then render a fixed output sequence, an assessment one-liner assembled by a template-based natural-language generator from the structured state, a prominent impression composed by deduplicating and ordering the fired rules' impression fragments, a synthesized sub-ten-item recommendation list, and an evidence-to-action section that exposes the full problem list, the per-problem detail with sources, and a finding-to-inference diagnostic trace.
All thresholds are isolated in one tunable constants block with provenance tags; abnormal physiology is reflected back through value-level color coding; and downstream artifacts (an SBAR escalation draft and a reveal-in-page SOAP note that marks unentered fields as unknown to double as a completeness tracker) are generated from the same state object, so every surface of the tool is driven by one source of truth and the decision path remains free of any probabilistic component.
Testing & Evaluation
-
AgentEVD is a deterministic, rule-based system, not a statistical model, so its evaluation is framed as specification-conformance and guideline-concordance rather than predictive accuracy on held-out data.
Each test verifies that an encoded rule fires exactly as specified, at the correct threshold and severity, with the correct source attribution, and that multi-rule interactions resolve, rank, and synthesize as designed.
The safety-critical property under test is that clinically dangerous states are never under-triaged: a false "no concern" is treated as the worst error class, and severity is designed to escalate, not relax, under uncertainty
-
Unit testing isolated individual rules in AgentEVD's rule set by entering only the minimum state needed to trip one predicate, confirming each fired alone, at the tier its severity function specifies, citing the source attached to that rule.
The cases span the engine's distinct rule families, the herniation red-flag, the BTF-anchored intracranial-hypertension rule, the IDSA-anchored infection rule, the decreased-drainage rule, the bradycardia/Cushing rule grounded in the Chowdhury review, and the hyponatremia rule, so that every independent branch of the logic is exercised once.
Two cases specifically tested AgentEVD's dynamic-severity functions rather than static firing: the decreased-drainage rule was confirmed to upgrade from INTERVENE to ESCALATE when high ICP was added, and the bradycardia rule was confirmed to render at ESCALATE with Cushing-physiology framing when paired with hypertension and to fall back when the blood pressure was normalized.
These confirm that AgentEVD's severity is computed from the combined state, not assigned per field.
The complete unit testing data can be viewed here:
-
System testing entered complete simulated patients to confirm that AgentEVD's engine integrates multiple concurrently firing rules the way its run-loop specifies: firing every matching rule, sorting them by severity, composing the impression by deduplicating and ordering the fired rules' impression fragments (including the postoperative qualifier), and collapsing the union of recommendations into the synthesized sub-ten-item list. The vasospasm-window patient and the rebleed-after-tumor-resection patient verified this on multi-concern hemorrhage-and-ICP presentations, confirming correct co-firing, ranking, and the postoperative-qualified impression. The infection-predominant patient with a reassuring exam was included specifically to test AgentEVD's restraint: it confirmed that when only the IDSA infection rule's conditions are met and no escalation predicate is satisfied, the tool leads with infection, surfaces dwell-time in the rationale, and does not manufacture a false escalation, demonstrating that the severity sort distinguishes a genuine emergency from a workup-level concern.
The complete system testing data can be viewed here:
-
Edge testing targeted AgentEVD's guardrails and boundaries rather than typical inputs.
The insufficient-data case confirmed the hard gate that blocks any recommendation until primary pathology and drain status are present, returning the explicit insufficient-data state instead of a guess.
The all-normal case confirmed the engine's no-rule-fired path resolves to routine surveillance rather than defaulting to alarm.
The ICP boundary case confirmed threshold conformance at the value written into the predicate.
The unknown-exam case confirmed two coupled guardrails fired together: the exam-reliability caveat printed and the optimism cap held, so an elevated-ICP state with an unassessable, sedated exam was not rendered as reassuring.
The conflicting-input case (high hourly output with a normal 24-hour total) confirmed that AgentEVD surfaces and ranks the relevant concern by severity rather than letting one field silently suppress another, validating that the rule predicates evaluate independently against the shared state.
The complete edge case testing data can be viewed here:
-
Bias testing exploited two structural facts about AgentEVD's engine: its firing predicates key only on clinical variables, never on protected demographic attributes, and missing data from a given source is handled as the absence of an abnormal finding from that source rather than as a penalty against the patient.
Two distinct invariance tests were run. The first held a complete clinical patient constant and varied only the sex field, confirming the assessment, impression, fired concerns, and recommendations were identical across variants, demonstrating that
AgentEVD cannot differentially triage on sex because sex is not an input to any predicate.
The second tested for technology-access bias, an equity dimension specific to neurocritical care, where access to CT/MRI and to invasive ICP monitoring is unevenly distributed across hospitals, payers, and levels of care: an infection-predominant patient was run with imaging omitted, and again with the ICP value omitted, and in both cases the output matched the run in which those sources were explicitly normal, confirming that
AgentEVD's infection pathway fires on its clinical and CSF criteria alone and neither suppresses nor downgrades a concern when an expensive or invasive test is simply unavailable.
The honest limitation stated alongside the clean results is that this design-level invariance does not remove guideline-level bias: AgentEVD's encoded thresholds inherit whatever population bias exists in the source guidelines, and the clinical values a user enters can themselves be measured or charted with bias the tool cannot detect or correct.
The complete bias testing data can be viewed here:
-
Usability was evaluated with the System Usability Scale (SUS), a validated ten-item questionnaire, administered to three users drawn from the tool's intended clinical audience.
Before rating, each user completed three standardized tasks on the live tool so that responses reflected hands-on use rather than first impression:
(1) enter a defined patient, a 60-year-old with aneurysmal subarachnoid hemorrhage, EVD open, highest ICP 24 mmHg, bloody CSF marked as a new change, and an examination unable to be assessed, then produce an assessment
(2) locate the field for a change in CSF appearance and record it; and (3) produce the SOAP note and copy it.
Each user then rated the ten SUS statements on a 1-to-5 scale (1 = strongly disagree, 5 = strongly agree).
The ten SUS items.
1. I would like to use this system frequently.
2. I found the system unnecessarily complex.
3. I thought the system was easy to use.
4. I think I would need technical support to use this system.
5. I found the functions well integrated.
6. I thought there was too much inconsistency.
7. I would imagine most people would learn this very quickly.
8. I found the system cumbersome to use.
9. I felt very confident using the system.
10. I needed to learn a lot before I could get going.
How the score was calculated:
SUS items alternate between positively and negatively worded statements, so raw scores are normalized before summing.
For the odd-numbered (positively worded) items 1, 3, 5, 7, and 9, the contribution is the raw score minus 1.
For the even-numbered (negatively worded) items 2, 4, 6, 8, and 10, the contribution is 5 minus the raw score.
This converts every item to a 0-to-4 scale where higher always means better usability.
The ten converted values are summed (range 0 to 40) and multiplied by 2.5 to yield a final score from 0 to 100.
The score is computed per respondent and then averaged.
SUS scores are interpreted against the established cross-study average of approximately 68; a mean above roughly 80 corresponds to the "excellent," A-range band.
Results
The three reviewers scored 80.0, 72.5, and 95.0, for an overall mean SUS of 82.5.
Examined by item category, the positively worded items (1, 3, 5, 7, 9, covering perceived ease, integration, learnability, and confidence) averaged a converted 3.5 out of 4 across reviewers, and the negatively worded items (2, 4, 6, 8, 10, covering complexity, need for support, inconsistency, and cumbersomeness) averaged a converted 3.1 out of 4; the slightly lower negative-item average was driven almost entirely by item 8 (cumbersome to use), where all three reviewers converged on the same mid-scale response, indicating the dense single-screen intake is the tool's main usability cost.
No reviewer flagged a need for technical support (item 4 converted to 4, 3, and 4), and all three rated integration of functions at the maximum (item 5 converted to 4 across the board), indicating the multi-section workflow held together coherently in use.
Interpreted against the established cross-study SUS average of approximately 68, all three individual scores were above average and the mean of 82.5 falls within the "excellent," A-range band.
This tells us the interface was learnable and usable by its intended clinical audience without training, that confidence and perceived ease were high, and that the one consistent friction point is the density of the single-screen form rather than any breakdown in the workflow or a need for external support. The spread from 72.5 to 95.0 reflects genuine variation in comfort with that density rather than uniform endorsement, which supports the credibility of the result and points directly to intake streamlining as the highest-value usability improvement.
The complete usability testing data can be viewed here:
-
Structured clinical feedback was gathered from three peer reviewers, who reviewed the live tool and responded to a focused set of questions on clinical logic, threshold appropriateness, completeness, escalation calibration, and floor usability.
Feedback was overwhelmingly positive on the tool's transparency, clinical tailoring, and documentation outputs, with a single consistent reservation about the time and familiarity required given the volume of information collected and produced.
The most frequently praised attribute was transparency and traceability. One reviewer specifically contrasted AgentEVD with conventional AI models, noting that it avoids the black-box problem because the problem list, the recommendations, and each output element are individually cited and exposed through the evidence-to-action framework, so a user can follow exactly how the tool reached its conclusion.
This is a meaningful endorsement of the core design decision to build a deterministic, fully attributed rule engine rather than a generative model, and it validates that the auditability goal was achieved in the eyes of clinical users.
Reviewers also praised the clinical relevance and the documentation outputs.
One noted that the intake was clinically relevant and tailored specifically to neurosurgery, and that the tool effectively doubles as a SOAP note and a rounding-report or rounding-list writer, with particular appreciation for its explicit documentation of pertinent negatives and pertinent unknowns.
Another highlighted the breadth and quality of the outputs, the assessment, the SBAR, the SOAP note, and the evidence framework, as a strength.
Across reviewers, the SOAP and SBAR generation and the structured problem list were repeatedly singled out as the most valuable features.
The one recurring critique concerned time and learning curve rather than correctness.
Two reviewers independently observed that, because of the amount of information the tool collects and outputs, it could be time-consuming in a fast-paced ICU or rounding setting, and that a clinician would need some time using it before feeling fully comfortable relying on it clinically.
Notably, neither framed this as a flaw in the clinical logic, the thresholds, or the outputs themselves; rather, it is a tradeoff inherent to the tool's thoroughness, and it points to the same intake-density issue surfaced independently in usability testing.
The complete peer feedback data can be viewed here:
Present and Future of
-
AgentEVD's central strength, confirmed independently by peer review, is transparency and traceability.
Every concern the engine raises is exposed through the evidence-to-action framework, where the problem list, the rationale, and each recommendation are individually attributed to a named source and evidence tier, so a user can follow the exact path from entered data to conclusion.
One reviewer explicitly contrasted this with conventional black-box AI models, identifying the fully auditable reasoning as the tool's distinguishing feature, which validates the foundational decision to build a deterministic, fully attributed rule engine with no language model in the decision path.
The same architecture produces a second strength, reproducibility: because the logic is specified rather than generated, identical inputs always yield identical outputs, a property directly verified across the unit, system, edge, and bias tests.
A third strength is clinical tailoring and documentation utility; reviewers praised the neurosurgery-specific intake and noted that the tool functions simultaneously as a decision-support aid and as a SOAP and SBAR generator, with explicit documentation of pertinent negatives and pertinent unknowns, making it useful as a rounding and handoff artifact rather than only an alerting tool.
Usability testing supports this, with a mean SUS of 82.5 placing the interface in the excellent band and all three raters above the cross-study average, and confidence and integration items scoring at or near the maximum.
Finally, the engine demonstrated correct contextual reasoning rather than rigid pattern-matching: dynamic-severity behavior was verified, and the bias testing surfaced a genuine equity property, namely that the absence of imaging or invasive ICP data is treated as the absence of an abnormal finding rather than as a penalty, so the tool does not disadvantage patients with less access to expensive monitoring.
-
The most consistent weakness, surfaced independently by both usability testing and peer feedback, is the cost of the tool's own thoroughness: the volume of information collected and produced makes the single-screen intake dense and potentially time-consuming in a fast-paced ICU or rounding environment.
Two peer reviewers noted they would need time with the tool before relying on it comfortably, and the SUS item on cumbersomeness was the single convergent friction point across all three raters, confirming that intake density, rather than any flaw in the clinical logic, is the primary usability limitation.
A second category of weakness is intrinsic to the design rather than a defect in execution: AgentEVD's knowledge is a fixed, hand-encoded corpus, so its recommendations are only as current and as complete as the encoded rules, it does not learn or update automatically, and it inherits whatever population bias exists in its source guidelines, a limitation the bias testing explicitly acknowledged.
The numeric thresholds, while provenance-tagged, include institutional conventions (for example certain output volumes) that are not themselves guideline-fixed and would require local calibration before any real use.
The evaluation itself carries methodological limits worth stating plainly: testing established specification-conformance and guideline-concordance on investigator-constructed cases rather than performance on real, prospectively collected patients; the usability and peer samples were small (three each); and no external or independent clinical validation has been performed.
Finally, the tool is deliberately narrow, adult-only and confined to the indwelling management phase, so it does not address insertion, pediatric patients, or many adjacent neurocritical-care decisions, and it remains an educational prototype explicitly not cleared for independent clinical use.
-
The highest-value improvement, pointed to by the convergence of the SUS and peer findings, is streamlining the intake without sacrificing the completeness reviewers valued.
A practical path is an abbreviated or express mode that surfaces only the fields relevant to the entered pathology and presenting concern, with the full intake available on demand, preserving the thoroughness as an option rather than a requirement.
A brief guided walkthrough or first-use onboarding would address the learning-curve comment directly, since reviewers framed their hesitation as familiarity rather than distrust.
On the knowledge side, the institutional-convention thresholds should be made explicitly user-configurable so a site can calibrate them to local protocol, converting a current limitation into a deliberate localization feature, and the source corpus should be placed on a defined review cycle so encoded rules are refreshed as guidelines are updated.
The evaluation should be strengthened by expanding case coverage to a larger and, ideally, externally sourced vignette set, and by enlarging the usability and peer samples to support firmer claims.
Adding a structured mechanism to capture which recommendations users accept, override, or find unhelpful would create a feedback loop that informs rule refinement over time.
-
The natural next step beyond this prototype is formal clinical validation: a blinded study in which expert raters score AgentEVD's outputs against an independent reference standard across a library of real, de-identified cases, measuring guideline concordance and, critically, escalation sensitivity, since a missed escalation is the worst error class for a tool of this kind.
That build-then-evaluate pathway was the explicit rationale for choosing a deterministic, auditable architecture, and it is what would distinguish this work from the many descriptive clinical chatbots in the literature.
Subsequent development could broaden scope in controlled increments, for example to EVD insertion-phase decision support or to pediatric populations with a separately validated rule set and thresholds, each treated as a distinct validation effort rather than an assumed extension.
The documentation strength reviewers identified suggests a second product direction: maturing the SOAP and SBAR generation into a fuller neurosurgical rounding and handoff tool, potentially with optional, clearly walled-off language-model assistance for prose generation that never touches the decision path, preserving the privacy and auditability guarantees.
Longer term, the configurable-threshold and feedback-capture features above would enable multi-site deployment in which each institution localizes the logic and contributes usage data, turning a single static prototype into a maintained, evidence-tracked platform, while retaining the founding principles that make it defensible: client-side privacy, deterministic reproducibility, and fully cited reasoning.
References
Carney N, Totten AM, O'Reilly C, et al. (2017). Guidelines for the management of severe traumatic brain injury, fourth edition. Neurosurgery, 80(1), 6–15. https://doi.org/10.1227/NEU.0000000000001432
Chowdhury T, Petropolis A, Cappellani RB. (2015). Cardiac emergencies in neurosurgical patients. BioMed Research International, 2015, 751320. https://doi.org/10.1155/2015/751320
Ebel F, Lichter E, Mariani L, Guzman R, Soleman J. (2023). Rapid versus gradual weaning of external ventricular drain: a systematic literature review and meta-analysis. Neurocritical Care, 39(1), 250–259. https://doi.org/10.1007/s12028-023-01766-6
Fried HI, Nathan BR, Rowe AS, et al. (2016). The insertion and management of external ventricular drains: an evidence-based consensus statement. Neurocritical Care, 24(1), 61–81. https://doi.org/10.1007/s12028-015-0224-8
Greenberg SM, Ziai WC, Cordonnier C, et al. (2022). 2022 guideline for the management of patients with spontaneous intracerebral hemorrhage: a guideline from the American Heart Association/American Stroke Association. Stroke, 53(7), e282–e361. https://doi.org/10.1161/STR.0000000000000407
Hoh BL, Ko NU, Amin-Hanjani S, et al. (2023). 2023 guideline for the management of patients with aneurysmal subarachnoid hemorrhage: a guideline from the American Heart Association/American Stroke Association. Stroke, 54(7), e314–e370. https://doi.org/10.1161/STR.0000000000000436
Meyrat R, Vivian E, Dulaney B, Barrera Gutierrez JC. (2024). Predictors of tract hemorrhage after external ventricular drain placement: a single-center retrospective study. Scientific Reports, 14, 27772. https://doi.org/10.1038/s41598-024-79421-z
Palasz J, D'Antona L, Farrell S, Elborady MA, Watkins LD, Toma AK. (2021). External ventricular drain management in subarachnoid haemorrhage: a systematic review and meta-analysis. Neurosurgical Review, 45(1), 365–373. https://doi.org/10.1007/s10143-021-01627-w
Tunkel AR, Hasbun R, Bhattacharjee A, et al. (2017). 2017 Infectious Diseases Society of America's clinical practice guidelines for healthcare-associated ventriculitis and meningitis. Clinical Infectious Diseases, 64(6), e34–e65. https://doi.org/10.1093/cid/ciw861