Voxa AI: Intraoperative Neuromonitoring Workflow Improvement

Inspiration

: Voxa AI was born from a critical gap in a proper training: the lack of standardized, structured educational platforms for new Intraoperative Neuromonitoring (IONM) technologists. Most of IONM training curriculum are heavily siloed within individual organizations, creating a unified standard across different service providers is incredibly impractical. Early on, we focused on internal educational initiatives, but faced three major roadblocks: The initial platforms were far too dense with theory for new technologists Building standalone machine learning models for each distinct IONM modality wasn’t viable as IONM sits at a complex intersection of neurosurgery, orthopaedics, anaesthesia, neurophysiology, and physics. A hardcoded decision-tree chatbot served as a helpful initial step, but it lacked the depth and adaptability for case variants. In addition, case reports are typically completed retrospectively, at the end of the case, rather than in real time. Technologists commonly rely on manual workarounds — using tools like the Windows Snipping Tool to capture waveform screenshots throughout the case — and then paste these into a report template in MS Word. This process is time-consuming, prone to inconsistency in what gets captured and labeled, and delays the availability of a finalized report. It also creates unnecessary cognitive load during a case, as technologists must remember to capture key moments manually rather than relying on the system to log them automatically. The Journey and Experimentation We turned to GPT vision to bridge this gap. Early tests were highly promising. The model successfully extracted crucial clinical values like latency, amplitude, and calibration data directly from SSEP windows. However, we encountered an important limitation. In certain MEP windows, latency or amplitude values aren’t explicitly shown on-screen. A human technologist has to manually calculate amplitude based on divisional counts and calibration data. While GPT vision could describe waveform characteristics well, it struggled to quantify these hidden amplitudes on its own. We also looked into using ChatGPT 5.6 to give a structured output based on the key anchors / markers that are tagged to the image, with acceptance from the technologist. The core distinction in displayed value extraction over measuring unlabeled waveform-shaped the system and to streamline workflow from monitoring start to an eventual automated report was what we built during Build Week. What We Built During Build Week, we connected our existing case-tracking and annotation components into a human-supervised Case Intelligence workflow. After the technologist permits screen capture of uploads screenshot, Voxa uses a general-purpose vision model to: Identify visible waveform changes State uncertainty when information is missing or ambiguous Extract visible labels and calibration information Recommend appropriate verification steps The technologist can seamlessly accept, correct or reject the model’s review or route the data to Sammy Annotation Lab for expert labelling and correction. This continuous feedback loop ensures we are actively building a high-quality dataset to lay the foundation for future, modality-specific computer vision models. Within the same interface, technologists can tag images with preset markers and log specific clinical events directly onto the waveform capture. At the conclusion of a case, these curated screenshots—complete with their associated events and markers—are automatically compiled into a comprehensive PDF case report. Consequently, the Voxa ecosystem provides a unified, end-to-end workflow consisting of: Waveform Review [Case Intelligence] Case Tracking [Case Intelligence] Report Generator [Case Intelligence] Annotation Lab [Sammy Annotation Lab] Training Modules [Learning Hub] Challenges We Ran Into / What We Learnt Before Build Week, we experimented with automatically placing measurement markers on waveform peaks and troughs. Although the system could generate markers, it could not consistently anchor them to the correct locations. Markers sometimes appeared on grid lines, channel dividers or other interface elements instead of the waveform. We also tested segmenting, skeletonizing and reconstructing waveform traces from screenshot pixels. Thin, partially hidden and overlapping traces made continuity tracing indeterminate. When two waveforms crossed, it could not reliably determine ownership of pixels to baseline or current trace or be able to reconstruct the obscured signal. A visually plausible reconstruction was not necessarily physiologically accurate. These experiments revealed a limitation in which screenshots are not equivalent to raw signal data and information on scale calibration, trace separation and waveform continuity may not be recoverable from pixels alone. Hence, during Build Week, we build based on the approach that AI should present visible and state uncertainty for human review, rather than claim measurements unsupported. Accomplishments Transformed earlier technical failures into a safer, human-supervised AI workflow Streamlined case tracking and annotation tools into one end-to-end process Established a practical foundation for collecting expert-reviewed data for future modality-specific models What’s Next for Voxa AI Build and train modality-specific computer vision models for waveform detection and measurement Validate the system’s performance with IONM technologists and clinical experts