Case Study: KBO Systems

Referenced research publication here:

Data Driven Investigation of Bispectral Index Algorithm (Nature)

1. Introduction

Overview:

KBO Systems is a pioneering medical technology startup focusing on advanced patient monitoring solutions. Their core objective was to create an algorithm capable of more accurately measuring patient sedation depth—a critical factor in anesthesia management and surgical procedures. Innovating upon the principles established by Medtronic’s BIS (Bispectral Index) Monitor, KBO Systems sought to develop a solution that leverages big data and machine learning to reduce the ambiguity and error sometimes seen in conventional sedation metrics.

Challenge:

While the BIS monitor is well-known and widely used, its proprietary algorithm is not fully disclosed. This opacity creates hurdles for practitioners who must interpret unusual BIS values influenced by factors other than actual sedation levels, such as EMG interference or certain pharmacologic agents. KBO Systems aimed to improve upon this existing technology by using big data and modern techniques (e.g., neural networks) to gain insights into the BIS calculation method and produce a more robust sedation index.

KBO Systems partnered with Benmore Technologies to replicate and refine the algorithm's functionality. Using publicly available datasets and advanced machine learning methods, Benmore Technologies engaged in a focused project to understand, improve, and patent an innovative sedation-depth measurement algorithm.

Objective:

• Utilize the VitalDB public dataset and other available EEG data to investigate the BIS algorithm and refine a neural network model for sedation depth measurement.
• Provide a technical foundation and supporting research that allows KBO Systems to file a provisional patent for their improved sedation depth algorithm.
• Conduct on-site testing of prototype devices to collect clinical data and further enhance the algorithm’s accuracy and reliability.

2. The Problem

Background:

Traditional BIS values correlate with anesthesia depth but can be influenced by a range of confounding factors—muscle activity, thermal warming devices, pacemakers, and certain drugs. This complexity often makes it challenging for anesthesiologists to distinguish between a genuinely light plane of anesthesia and false elevations caused by extraneous signals.

The paper "Data Driven Investigation of Bispectral Index Algorithm" (Scientific Reports, 2019) highlights an approach to understanding the BIS calculation from EEG-based subparameters (BSR, EMG, SEF, RBR) using clinical big data and machine learning. The study demonstrated that the BIS algorithm likely employs multiple regression equations with different weights depending on the level of sedation, but key parts of the proprietary algorithm remain unknown.

Pain Points:

• Opacity of BIS Algorithm: Without full disclosure of how the BIS is calculated, interpreting unusual readings is difficult.
• Artifact Intrusion: EMG signals and other environmental factors can skew BIS values, leading to clinical misinterpretation.
• Limited Insight into Proprietary Metrics: While publicly known subparameters (BSR, SEF, EMG, RBR) exist, some remain undisclosed (e.g., QUAZI suppression index), limiting replicability and improvement.

3. Our Solution

Discovery Process:

Benmore Technologies began by reviewing the Nature publication and its methodology for dissecting the BIS algorithm. This gave us insight into the use of big data and machine learning for unraveling parts of the BIS calculation. We obtained the VitalDB public dataset—comprising EEG data and corresponding BIS values from thousands of surgeries—and processed it to filter high-quality, artifact-free EEG segments.

From there, we employed a custom neural network approach. Our model was designed to:

1. Integrate known EEG subparameters (BSR, SEF, EMG, RBR).
2. Infer missing components through pattern recognition in the large dataset.
3. Predict BIS-like sedation depth scores with improved robustness against known artifacts and confounding factors.

Proposed Solution:

• Neural Network Architecture: A deep learning model trained on the VitalDB data to predict a “BIS-like” score.
• Feature Refinement: Incorporation of established subparameters (BSR, EMG, SEF, RBR) alongside additional spectral and bispectral EEG features derived from the raw signals.
• Iterative Enhancement: Continuous testing and refinement using both the VitalDB dataset and on-site clinical data collection. The on-site clinical tests provided a feedback loop to verify the algorithm’s performance in real-world conditions.

Technology Stack:

• Data Processing: Python (NumPy, Pandas) for data wrangling and filtering EEG signals.
• Modeling & Analysis: TensorFlow/Keras for training neural networks and scikit-learn for complementary machine learning tasks.
• Infrastructure: Cloud-based GPU computing to handle large-scale EEG data processing and model training efficiently.

4. Implementation

Challenges Encountered:

• Complexity of EEG Data: EEG data can be noisy and susceptible to various artifacts. It was critical to filter and pre-process the data rigorously before training the neural network.
• Interpretability of Results: While the neural network improved on raw correlation with sedation depth, ensuring that the model’s predictions were clinically interpretable required careful validation.

Timeline:

1. Data Preparation (2–3 Months): Gathered the VitalDB dataset and applied data quality filters, ensuring SQI > 90%.
2. Model Development (3–4 Months): Developed and refined neural network models, iteratively testing performance on the training and validation sets.

Collaboration:

Benmore Technologies worked closely with KBO Systems to ensure alignment of technical goals with clinical needs. Through regular video conferences and onsite visits, the joint team streamlined data collection, model training, and validation efforts.

5. Results

Key Outcomes:

• Provisional Patent Filed: Leveraging our research and initial neural network modeling, KBO Systems filed a provisional patent for their innovative sedation depth algorithm. This intellectual property milestone underpins their competitive advantage.
• Refined Sedation Metric: The resulting model showed promising consistency in correlating EEG patterns with sedation levels.
• Investor Confidence & Funding: With a solid technological foundation and patent protection in place, KBO Systems successfully raised a seed round. The funding will support further R&D, regulatory pathways, and clinical trials.

6. Lessons Learned

Navigating Proprietary Systems:

Working around a proprietary, non-disclosed algorithm required a data-driven reverse engineering approach. While complete replication of the BIS algorithm remains challenging due to unknown factors (like QUAZI suppression index calculation), we successfully replicated the general algorithm with the developed model.

7. Conclusion

Summary:

Through an intensive research and development effort, Benmore Technologies assisted KBO Systems in dissecting aspects of the BIS algorithm and creating a novel sedation depth measurement model. By leveraging big data, machine learning, and onsite clinical validation, we helped lay the groundwork for a new generation of anesthesia monitoring tools.

Call to Action:

If you require assistance in medical device algorithm development, data analysis, or technical validation for your innovation, please don’t hesitate to book a call with us! Click the "Let's Talk" button.