The research team, led by investigators at the University of Hong Kong and Ningbo Hospital of Integrated Traditional Chinese and Western Medicine in China, took a deliberately modular approach. Rather than asking an AI to produce an end-to-end diagnosis from an image, they split the problem into two separate modules. The first tested whether a large multimodal model could transcribe and interpret pure-tone audiometry and tympanometry images. The second tested whether AI workflows could calculate twenty-five prespecified clinical fields and draft professional and patient-facing reports in Chinese, given already-verified structured data. This separation matters because a single end-to-end score can hide whether errors come from reading the image, applying specialty rules, or communicating the result.

The study drew on 158 outpatient audiology encounters collected between December 2024 and March 2025, of which 155 records representing 151 unique patients and 302 ears were eligible. Reference standards were built painstakingly: two trained transcribers independently entered every threshold while masked to each other, and two audiologists with 13 and 16 years of clinical experience classified tympanogram curves, agreeing on 296 of 300 dually classified ears, a Cohen’s kappa of 0.979. Air-bone gaps, hearing-loss degrees, and loss types were derived under prespecified rules, including a local convention that a meaningful air-bone gap required at least two comparable frequencies with gaps of 15 dB or more.

The heart of the first module was the audiologist skill: a carefully engineered prompt, not a fine-tuned model, that encoded audiological practice. Early development errors were revealing. The model initially produced thresholds not on the standard 5-dB grid, misapplied degree boundaries, overcalled conductive components, treated insufficient bone-conduction evidence as a negative air-bone gap, and read cancelled 95-dB acoustic-reflex marks as present responses. The refined skill imposed a fixed sequence: verify the image and ear, inspect axes and legends, assign symbols, transcribe before calculating, preserve no-response entries, then derive and cross-check. When masking could not be assigned unambiguously, the prompt instructed the model to abstain rather than guess.

The results of the locked test were striking. In 50 development records, the skill raised hearing-loss-type agreement from 85.0 to 94.0 percent and acoustic-reflex agreement from 70.4 to 98.4 percent compared with a schema-only prompt. In the held-out evaluation of 101 independent patients using a Codex GPT5.5 agentic workflow, the system achieved 1809 of 1820 exact numeric thresholds, 99.4 percent, and 91 of 101 patients met every numeric and no-response criterion. Degree was correct in all 101 patients, hearing-loss type in 100, and tympanometry measurements in all 578 entries. Yet the Achilles’ heel persisted: only 9 of 15 no-response entries were correct, and 18 of 19 air-bone-gap mismatches occurred because the model forced a negative label when the evidence supported an indeterminate one.

A post hoc robustness analysis using DeepSeek-V4-Flash-Vision-Exp on the same patients showed how much performance depends on the implementation. With the same locked skill, hearing-loss-type agreement rose from 37.6 to 70.3 percent, a dramatic improvement, but exact numeric-threshold agreement reached only 72.1 percent, reflex agreement hovered near 69 percent, and no-response agreement was zero. The authors are careful to note this was a descriptive cross-model comparison, not a matched foundation-model experiment, since the execution environments differed. The lesson, however, is clear: specialty guidance can substantially improve rule-dependent interpretation, but raw visual accuracy remains tied to the underlying model and workflow.

The second module addressed reporting. Here the AI received adjudicated structured values rather than its own image predictions, calculated twenty-five prespecified fields, and drafted separate Chinese professional and patient-facing reports. After two audiologists reviewed 60 initial cases and identified overreliance on the speech-frequency average, omission of high-frequency losses, and weak integration of history with results, the prompts were refined and locked. In the formal evaluation of 91 independent patients, all twenty-five rule-derived fields were correct in 91 of 91 Codex-workflow cases and 87 of 91 DeepSeek-workflow cases. Both audiologists rated every single report from both workflows as accurate or basically accurate; no report received a rating of clear error or potentially misleading.

An exploratory lay evaluation added a human dimension. Five lay raters compared pre-refinement patient-facing reports from the two workflows in 48 patients. DeepSeek reports were preferred for explanations in 44 of 48 patients and for next steps in 30, but they were also significantly longer, with a median of 392 versus 235 Chinese characters. Overall preference and perceived ease did not differ significantly. The authors emphasize that preference and readability proxies do not establish comprehension, and that no lay evaluation of the refined reports was conducted. This matters because hearing-health materials often exceed recommended reading levels, and limited health literacy can coexist with hearing loss in older adults.

The study’s limitations are candidly enumerated. It came from a single hospital with two devices over four months. Only 15 no-response entries existed, from just three validation patients. The two audiologists who rated the final reports were the same ones who had refined the prompts, raising the possibility of incorporation bias. Even with zero unfavorable ratings among 91 patients, the statistical upper bound on the unfavorable-report rate remains roughly 4.1 percent. Reproducibility was constrained by reliance on proprietary services: the exact Codex snapshot and sampling settings were unavailable, and provider data retention could not be excluded. No end-to-end test connected the two modules, so error propagation from image to report was never measured.

What the study ultimately offers is an architecture rather than a product. The authors envision a safety-conscious pipeline in which image transcription, deterministic validation and calculation, report drafting, uncertainty flags, and clinician approval remain visible, auditable handoff points. This aligns with what Chinese audiologists themselves have said in qualitative work: AI may absorb repetitive technical work, but communication, judgment, and responsibility should remain clinician-led. The findings support prospective evaluation of modular, clinician-supervised assistance, not autonomous diagnosis. The next step, the authors argue, is a silent prospective deployment that connects the modules while preserving intermediate outputs, measuring abstention, correction burden, review time, patient comprehension, and downstream clinical decisions. Until then, the audiogram-reading AI remains a promising apprentice, one that can transcribe nearly every threshold perfectly yet still needs its audiologist to teach it what silence means.

Subject of Research: Audiologist-guided multimodal AI for interpreting pure-tone audiometry and tympanometry and generating clinical reports

Article Title: Audiologist-Guided Multimodal AI for Pure-Tone Audiometry and Tympanometry Interpretation and Reporting

Article References: Wu, X., Shen, X., Mo, C., Shao, S., Wang, J., & Wang, S. (2026). Audiologist-Guided Multimodal AI for Pure-Tone Audiometry and Tympanometry Interpretation and Reporting. Journal of Medical Systems, 50(1), Article 140. https://doi.org/10.1007/s10916-026-02463-5

Image Credits: AI Generated

DOI: 10.1007/s10916-026-02463-5

Keywords: artificial intelligence, audiology, pure-tone audiometry, tympanometry, large language models, multimodal AI, clinical decision support, hearing loss, medical imaging AI, patient communication, prompt engineering, Journal of Medical Systems

Yushan Wei and Lien-Chung Wei, both of the Taoyuan Psychiatric Center in Taiwan, published the correspondence on 30 September 2026, drawing on frontline psychiatric nursing experience and a review of the emerging literature on ambient documentation. Their argument is deceptively simple. In psychiatry, the clinical record is not merely an administrative byproduct of care; it is itself a clinical instrument. Nurses on inpatient psychiatric units observe patients continuously across shifts, in moments when no physician is present, and those observations—sleep patterns, appetite, social withdrawal, agitation, subtle changes in speech or self-care—are often the earliest signals of deterioration, relapse, or risk. If ambient AI systems are designed, as they currently are, around the doctor-patient consultation as the canonical unit of documentation, they may systematically filter out exactly the information that psychiatric nursing contributes.

The technical architecture of ambient scribes helps explain the concern. These systems typically capture audio during a scheduled clinical encounter, apply automatic speech recognition, and then use large language models to organize the transcript into a structured note: history, examination, assessment, plan. The template is inherited from physician documentation norms, and the summarization step is trained to prioritize what a physician would conventionally record. A nurse’s longitudinal observations do not arrive as a discrete encounter with a clean audio capture. They accumulate across a shift, embedded in handover conversations, charting snippets, and informal exchanges. There is no microphone positioned to capture them, and even if there were, a summarization model optimized for consultation structure would have no obvious slot in which to place them.

The authors point to recent studies that have documented both the promise and the blind spots of these tools. A 2026 qualitative study in the same journal explored ambient AI for inpatient documentation with junior doctors and found enthusiasm for reduced administrative burden, but its focus remained squarely on physician workflows. Meanwhile, a study published in JMIR Nursing examined a nurse-led ambient AI scribe applied to patient safety incident investigation reports and found measurable improvements in document quality, suggesting that nurses can benefit from the technology when it is deliberately adapted to their tasks. A JAMA Psychiatry study of AI scribe use in psychiatric documentation in primary care likewise demonstrated feasibility in mental health settings. Yet, the correspondence argues, none of these lines of work addresses the specific epistemic role of inpatient psychiatric nursing observation, which is continuous rather than episodic and behavioral rather than dialogic.

This gap is not a minor design quirk. Prior research on nursing documentation has shown that poor or incomplete records are a genuine patient safety issue. A 2021 analysis in Frontiers in Computer Science identified barriers that healthcare professionals and students face in documenting nursing care, including time pressure, unclear standards, and electronic systems that were not built with nursing workflows in mind. The risk identified by Wei and Wei is that ambient AI could compound these barriers invisibly. A handwritten or free-text nursing note, however imperfect, at least exists as a space where observation can be recorded. If an AI-generated note becomes the dominant record and its template has no place for behavioral observation, the omission happens upstream, before any human decides what to write.

There is also a subtler danger: the illusion of completeness. Large language model summaries are fluent, well-organized, and confident in tone, which can make a note that omits nursing observations appear comprehensive rather than partial. Clinicians reading a polished AI-generated record may assume that everything salient has been captured, and may be less likely to consult separate nursing notes or to ask ward staff directly. In psychiatric care, where decisions about observation levels, leave privileges, and medication changes often hinge on nursing input, this smoothing effect could have concrete clinical consequences. The correspondence frames this as a problem of preservation: the goal is not to reject ambient AI but to ensure that the specific knowledge nurses produce survives the transition to automated documentation.

What would preservation look like in practice? The authors propose evaluation methods rather than a finished technical solution, reflecting the correspondence format. Ambient systems deployed on psychiatric wards should be explicitly tested for whether nursing observations are retained in the generated record, not just whether physician documentation improves. That means evaluation datasets and checklists that include nursing-specific content: sleep and activity patterns, eating behavior, medication adherence observed on the ward, social interaction, signs of agitation or withdrawal, and responses to nursing interventions. It also means involving psychiatric nurses in the design of templates and summarization prompts, so that the output schema has dedicated space for longitudinal behavioral observation rather than forcing everything into a consultation-shaped container.

The technical challenges are real but not insurmountable. Speech recognition on a noisy ward raises privacy and consent questions that are sharper in psychiatry than elsewhere, since patients may be acutely unwell and their capacity to consent to continuous recording may fluctuate. Audio capture of informal ward interactions would be legally and ethically fraught in most jurisdictions. A more realistic path may be hybrid: ambient AI handles the structured consultation note, while nurses use voice-dictated or AI-assisted entry modes tailored to observation charting, with the two streams merged into a single patient record that visibly distinguishes their sources. The nurse-led incident reporting study suggests that when the task is defined around nursing work, the technology can deliver quality gains; the lesson is that task definition, not the model, is the binding constraint.

The correspondence also arrives at a moment of institutional reckoning about AI in clinical records. The authors themselves disclose that OpenAI Codex was used for literature discovery, drafting, and revision of their manuscript, with both authors reviewing and taking responsibility for the final text—a transparency practice that mirrors the disclosure norms now expected of AI scribes in clinical settings. That symmetry is fitting. The central question they raise about psychiatric documentation is ultimately a question about any AI-mediated record: who decides what counts as clinically salient, and can the professions whose knowledge is least template-friendly push back before the defaults harden?

For psychiatric nursing, the stakes are unusually high because observation is the profession’s core diagnostic contribution. A patient who has stopped eating, a sudden shift from withdrawal to uncharacteristic cheerfulness that can precede a suicide attempt, the early tremor and restlessness of medication side effects—these are detected by nurses who spend hours with patients, not by a microphone that switches on when the psychiatrist enters the room. Wei and Wei’s intervention is a warning delivered early enough to matter: ambient AI on psychiatric wards should be evaluated not only by how well it writes what doctors say, but by whether it preserves what nurses see. If the technology is adapted with that standard in mind, it could lighten the documentation load across the whole multidisciplinary team. If it is not, hospitals may gain efficiency while quietly losing one of the oldest and most valuable instruments in mental health care: the trained, continuous, human eye of the ward nurse.

Subject of Research: Ambient AI clinical documentation and nursing observations in psychiatric inpatient care

Article Title: Preserving Nursing Observations in Ambient AI Documentation on Psychiatric Wards

Article References: Wei, Y., & Wei, L.-C. (2026). Preserving Nursing Observations in Ambient AI Documentation on Psychiatric Wards. Journal of Medical Systems, 50(1), Article 139. https://doi.org/10.1007/s10916-026-02467-1

Image Credits: AI Generated

DOI: 10.1007/s10916-026-02467-1

Keywords: ambient AI, AI scribes, psychiatric nursing, clinical documentation, nursing observations, electronic health records, patient safety, large language models, mental health care, health informatics, documentation quality, Journal of Medical Systems

The debate centers on a study by Backes and colleagues, published earlier in the same journal, which asked whether the choice of hospital information system influences digital maturity scores. Digital maturity, in this context, is typically measured through structured national assessment frameworks that grade hospitals on the sophistication of their clinical, administrative, and technical digital capabilities, from electronic documentation and data exchange to advanced analytics and telehealth integration. Because hospitals overwhelmingly rely on commercial vendors for their core information systems, the intuition that vendor characteristics, such as market presence, product breadth, or implementation experience, might correlate with maturity outcomes is compelling. Policymakers increasingly want to know whether choosing the right vendor can accelerate digital transformation, and vendor selection has become a strategic decision with multi-million-dollar consequences.

Lyu and colleagues do not dispute that the question matters. What they dispute is whether the analytical approach used to answer it can support the conclusions drawn. Their commentary organizes its critique around three technical pillars: provider-level inference, self-inclusion, and market-size dependence. Each of these describes a distinct way that the statistical machinery of the original analysis could produce misleading estimates, and together they form a checklist that the authors believe should be applied to any study attempting to connect vendor characteristics to hospital outcomes.

The first pillar concerns what the authors call provider-level inference. When researchers examine whether vendor characteristics are associated with hospital digital maturity, the vendor characteristics themselves vary at the level of the provider, not the hospital. Multiple hospitals in a dataset may share the same vendor, meaning their values on vendor-level explanatory variables are identical. Ignoring this clustering and treating each hospital as an independent observation inflates the effective sample size for vendor-level effects and can dramatically overstate statistical confidence. The correspondence points to the established literature on cluster-robust inference, including widely cited methodological guidance by MacKinnon, Nielsen, and Webb, which emphasizes that standard errors must account for the structure of the data. Recent work by Huang on the failure of cluster-robust methods in small samples adds a further caution: when the number of clusters, in this case the number of distinct vendors, is limited, even cluster-robust corrections can perform poorly, producing confidence intervals that are too narrow and p-values that are too optimistic.

The implications are substantial. If a study includes thousands of hospitals served by only a handful of major vendors, the effective information about vendor-level effects is bounded by the number of vendors, not the number of hospitals. Any claim that a particular vendor attribute, such as market share or product portfolio breadth, is significantly associated with maturity outcomes must survive inference procedures that respect this clustering structure. The commentary argues that without such corrections, the reported associations may reflect statistical artifacts rather than genuine market dynamics, and the field risks building policy recommendations on findings that would not replicate.

The second pillar, self-inclusion, addresses a subtler but equally consequential problem. In many studies of this type, the vendors being evaluated as potential drivers of digital maturity are themselves embedded in the market being studied, and in some analytical framings, entities can effectively appear on both sides of the regression equation. When a provider characteristic is derived from data that includes the very hospitals whose outcomes it is meant to predict, the explanatory variable and the outcome variable become mechanically entangled. This can induce spurious correlation: the predictor partially contains information about the outcome by construction, rather than by any real-world causal pathway. Lyu and colleagues argue that the original analysis did not adequately separate the measurement of provider characteristics from the hospital populations used to compute them, leaving open the possibility that at least part of the observed association is an artifact of this circularity rather than evidence that vendor choice shapes maturity.

The third pillar, market-size dependence, concerns how vendor characteristics are defined and scaled. Characteristics such as vendor market share are inherently relative quantities that depend on the size and composition of the market being measured. A vendor serving a large fraction of hospitals in one region may serve a tiny fraction in another, and the same vendor may occupy different market positions in different hospital segments. If the analysis pools heterogeneous markets or computes vendor characteristics over an ill-defined population, the resulting measures can conflate vendor quality or strategy with simple market structure. An association between market share and digital maturity might then reflect regional differences in healthcare infrastructure, funding, or policy environments, rather than any property of the vendors themselves. The commentary suggests that without careful attention to how the relevant market is delimited and how provider characteristics are normalized, the estimated relationships remain open to confounding by market size.

The correspondence also situates the debate within a broader international context. Assessing hospital digital maturity has become a priority across health systems, and a recent viewpoint in the Journal of Medical Internet Res compared national assessment approaches in five countries, revealing just how differently countries operationalize the concept. Meanwhile, surveys of digital health companies’ experiences with electronic health record interfaces, including work published in the Journal of the American Medical Informatics Association, highlight how deeply vendor capabilities and interoperability practices shape what hospitals can actually achieve with their systems. Taken together, this literature underscores that vendor-hospital relationships are real and consequential, which makes it all the more important, the authors contend, that the statistical evidence linking them be rigorous.

For hospital leaders and procurement officials, the practical message is one of caution. If the association between vendor characteristics and digital maturity is weaker or less certain than early studies suggest, then decisions driven by the assumption that a particular vendor guarantees maturity gains may be misplaced. Investments in organizational readiness, staff training, workflow redesign, and governance may matter as much as or more than vendor selection, a conclusion consistent with decades of health informatics research showing that technology adoption succeeds or fails on sociotechnical grounds. The commentary does not claim that vendors are irrelevant; rather, it insists that the field currently lacks the inferential rigor needed to quantify exactly how much vendor characteristics contribute to maturity outcomes.

The authors of the correspondence, who report no funding and no competing interests, frame their intervention as constructive: a set of analytical safeguards, provider-level clustering with appropriate robust or hierarchical standard errors, careful exclusion of self-referential constructs, and explicit attention to market definitions, that future studies should adopt before drawing policy-relevant conclusions. As health systems worldwide pour resources into digital transformation and vendors compete to position their platforms as engines of maturity, the message from Hangzhou is clear: before declaring that system choice drives digital maturity, researchers must first ensure their statistics can legitimately make that claim. Until then, the true drivers of hospital digitalization remain an open and urgently important question.

Subject of Research: Methodological critique of statistical associations between health information system provider characteristics and hospital digital maturity

Article Title: Clarifying Associations Between HIS Provider Characteristics and Hospital Digital Maturity: Provider-Level Inference, Self-Inclusion, and Market-Size Dependence

Article References: Lyu, H., Hu, Y., & Fan, S. (2026). Clarifying Associations Between HIS Provider Characteristics and Hospital Digital Maturity: Provider-Level Inference, Self-Inclusion, and Market-Size Dependence. Journal of Medical Systems, 50(1), Article 129. https://doi.org/10.1007/s10916-026-02456-4

Image Credits: AI Generated

DOI: 10.1007/s10916-026-02456-4

Keywords: hospital information systems, digital maturity, health IT vendors, cluster-robust inference, provider-level analysis, market share, electronic health records, health informatics, statistical methodology, hospital digitalization, vendor selection, Journal of Medical Systems