Derivative works — US Patent 11484284

Defensive Disclosure and Prior Art Derivations for US 11,484,284

Publication Date: May 1, 2026
Subject: Defensive publication of technical derivations and alternative embodiments related to US patent 11,484,284, "Methods and devices for processing heart sounds."
Purpose: To place in the public domain a series of technical disclosures that expand upon, vary, and apply the core concepts of US patent 11,484,284. This document is intended to serve as prior art against future patent applications for incremental or obvious improvements in the field of acoustic signal processing for physiological monitoring and automated therapy.

Derivations of Independent Claim 1: System for Processing Heart Sound Signals

Claim 1 of US 11,484,284 describes a system comprising a phonocardiogram recorder, processor, and display that measures, segments, and extracts acoustic features from heart sounds, using unsupervised machine learning to group the sounds by disease stage and recommend therapy. The following are derivative embodiments.

1. Material & Component Substitution

1.1. Piezoelectric MEMS Sensor Array for Acoustic Recording

Enabling Description: The phonocardiogram recording device is substituted with a flexible, conformable patch containing an array of Micro-Electro-Mechanical System (MEMS) piezoelectric acoustic sensors. Each sensor in the 2D array (e.g., 8x8) is fabricated on a silicon substrate with a PZT (lead zirconate titanate) thin film. The array provides spatial resolution across the chest, allowing the system to capture sounds from different valve locations simultaneously. A multiplexer (MUX) scans the array, and the signals are fed into a multi-channel analog-to-digital converter (ADC). The processor then uses beamforming algorithms to isolate sound sources, enhancing the signal-to-noise ratio for the target heart valve (e.g., aortic or mitral) and rejecting noise from other physiological sources like respiration or digestion. This component substitution improves signal source separation beyond what a single-point sensor can achieve.

Diagram:

graph TD
    subgraph Conformable Patch
        A[MEMS Sensor 1,1] --> MUX;
        B[MEMS Sensor 1,2] --> MUX;
        C[...] --> MUX;
        D[MEMS Sensor 8,8] --> MUX;
    end
    MUX --> ADC[Multi-Channel ADC];
    ADC --> Processor;
    Processor --> DSP{Digital Signal Processor};
    DSP -- Beamforming Algorithm --> Feature_Extraction[Feature Extraction Module];
    Feature_Extraction --> ML[Unsupervised ML Module];
    ML --> Display[Display/Therapy Recommendation];

1.2. Neuromorphic Spike-Based Processing Unit

Enabling Description: The general-purpose processor is replaced with a neuromorphic processing unit (NPU), such as an asynchronous spiking neural network (SNN) hardware accelerator (e.g., Intel Loihi or IBM TrueNorth architecture). The digitized audio signal from the ADC is first converted into a sequence of spikes using a delta modulator. The segmentation and feature extraction steps are implemented as layers within the SNN. The network learns to recognize temporal patterns corresponding to S1 and S2 sounds through spike-timing-dependent plasticity (STDP), an unsupervised learning rule. This approach significantly reduces power consumption, as computation is event-driven (i.e., occurs only when spikes are present), making it ideal for continuous, long-term monitoring in a battery-powered wearable device. The clustering of disease stages is an emergent property of the network's learned representations.

Diagram:

sequenceDiagram
    participant ADC
    participant DeltaMod as Delta Modulator
    participant NPU as Neuromorphic Processor (SNN)
    participant Memory as Machine-Readable Medium

    ADC->>DeltaMod: Digitized audio stream
    DeltaMod->>NPU: Asynchronous spike train
    NPU->>NPU: Process spikes via STDP
    Note right of NPU: Segmentation and feature extraction occur as emergent properties of the network's learned weights.
    NPU->>Memory: Store learned cluster centroids
    Memory-->>NPU: Load centroids for inference
    NPU->>Display: Output Disease Stage / Recommendation

2. Cross-Domain Application

2.1. Aerospace: Acoustic Monitoring of Turbine Blade Health

Enabling Description: The system is adapted for monitoring the health of jet engine turbine blades during operation. The phonocardiogram recorder is replaced by a set of high-temperature, wide-bandwidth piezoelectric sensors mounted on the engine casing. The "heart sound" is the acoustic signature of the rotating blades. The system segments the continuous acoustic signal based on blade passing frequency, isolating the signature of each blade. Acoustic features extracted include spectral kurtosis, dominant resonance frequencies, and power spectral density in ultrasonic bands. The unsupervised machine learning algorithm (e.g., k-means or DBSCAN) clusters these feature sets to identify deviations from a healthy blade signature, indicating the early formation of micro-cracks or material fatigue. This allows for predictive maintenance before a catastrophic failure occurs.

Diagram:

flowchart LR
    subgraph Engine Casing
        S1[Sensor 1]
        S2[Sensor 2]
        S3[Sensor N]
    end
    subgraph Processing Unit
        Processor -- Segmentation by Blade Freq. --> Segments[Blade Signatures]
        Segments -- Feature Extraction --> Features[Spectral Kurtosis, PSD, etc.]
        Features -- Unsupervised Clustering --> Clusters[Health States: Healthy, Warning, Critical]
        Clusters --> Maint[Maintenance Alert System]
    end
    S1 & S2 & S3 --> Processor

2.2. AgTech: Non-Invasive Bovine Cardiopulmonary Monitoring

Enabling Description: The device is reconfigured as a durable bolus that is ingested by cattle and resides in the reticulum, or as a ruggedized ear tag. The acoustic sensor is a robust, waterproof transducer. The system continuously monitors bovine heart and lung sounds. The unsupervised machine learning model is trained to distinguish between normal cardiac and respiratory sounds and signatures indicative of conditions like Bovine Respiratory Disease (BRD) or hardware disease. By analyzing trends in heart rate variability, S1/S2 amplitude ratios, and the presence of wheezing or crackling in the lung sounds, the system can provide early warning of illness for an individual animal, allowing for isolation and treatment before the disease spreads throughout the herd. Data is transmitted wirelessly via a LoRaWAN network to a central farm management system.

Diagram:

graph TD
    A[Ear Tag/Bolus Sensor] --> B{Signal Processing};
    B --> C[Feature Extraction (HRV, S1/S2 Ratio, Wheeze Freq)];
    C --> D{Unsupervised ML Classifier};
    D --> E{Health Status};
    E --> F[LoRaWAN TX];
    F --> G((Farm Management Dashboard));
    D -- Anomaly Detected --> E;
    subgraph On-Animal Device
        A
        B
        C
        D
        E
        F
    end

2.3. Industrial Manufacturing: Gearbox Fault Prediction

Enabling Description: The method is applied to predict failures in industrial gearboxes. A vibration sensor (accelerometer) and a high-frequency acoustic sensor are attached to the gearbox housing. The "cardiac cycle" is one full rotation of the primary gear shaft, with a tachometer providing the timing reference for segmentation. The system extracts acoustic and vibrational features such as the energy of gear mesh frequencies and their sidebands, as well as oil-film-induced acoustic emissions. The unsupervised machine learning algorithm clusters the operational states of the gearbox over time. A gradual shift in cluster centroids indicates progressive wear (e.g., pitting, spalling), while a sudden jump indicates an acute fault like a tooth breakage. This provides a real-time health assessment and enables condition-based maintenance.

Diagram:

stateDiagram-v2
    [*] --> Healthy
    Healthy --> GradualWear: Feature vector drifts
    GradualWear --> ImpendingFailure: Cluster centroid crosses threshold
    ImpendingFailure --> MaintenanceAlert
    Healthy --> AcuteFault: Sudden feature shift
    GradualWear --> AcuteFault: Sudden feature shift
    AcuteFault --> ShutdownAlert
    MaintenanceAlert --> [*]
    ShutdownAlert --> [*]

    note left of Healthy
      Acoustic features
      (e.g., gear mesh sidebands)
      are stable within a defined cluster.
    end note

3. Integration with Emerging Tech

3.1. Federated Learning for Privacy-Preserving Model Training

Enabling Description: To train the unsupervised machine learning model without centralizing sensitive patient heart sound data, a federated learning architecture is employed. Each wearable device (client) runs the segmentation and feature extraction steps locally. It then uses its local data to compute an update to a global clustering model (e.g., updated centroid positions for a k-means model). Only these non-identifiable model updates are encrypted and sent to a central server. The server aggregates the updates from many users (e.g., using Federated Averaging) to create an improved global model, which is then sent back to the clients. This process repeats, allowing the model to learn from a diverse population dataset while raw physiological data never leaves the user's device, thus preserving privacy.

Diagram:

sequenceDiagram
    participant Client1 as Wearable Device 1
    participant Client2 as Wearable Device 2
    participant Server

    loop Training Round
        Server-->>Client1: Send Global Model
        Server-->>Client2: Send Global Model
        Client1->>Client1: Compute model update on local data
        Client2->>Client2: Compute model update on local data
        Client1-->>Server: Send encrypted model update
        Client2-->>Server: Send encrypted model update
        Server->>Server: Aggregate updates (Federated Averaging)
        Server->>Server: Create new Global Model
    end

Derivations of Independent Claim 17: Wearable System with Automated Treatment

Claim 17 of US 11,484,284 details a specific wearable system with an integrated treatment device that administers a compound based on real-time analysis of acoustic features like the area under the curve (AUC) of the power spectral density (PSD).

1. Material & Component Substitution

1.1. Transdermal Microneedle Array for Drug Delivery

Enabling Description: The treatment system, described as a reservoir and administration device, is implemented as a disposable patch containing a microneedle array. The needles are fabricated from a biodegradable polymer (e.g., polylactic-co-glycolic acid, PLGA) and are loaded with the anti-remodeling therapy. The patch is communicatively coupled to the main processing unit. When the processor determines that the PSD's AUC has crossed a predetermined threshold, it sends a signal to an actuator in the patch. The actuator, which can be a micro-pump or a system that applies a small electrical current to trigger electro-responsive polymer needles, initiates the delivery of the drug transdermally. This provides a minimally invasive, painless, and controlled delivery mechanism compared to a conventional pump and catheter.

Diagram:

classDiagram
    class ProcessingUnit {
        +analyzeHeartSound()
        +calculatePsdAuc()
        +triggerDelivery()
    }
    class MicroneedlePatch {
        -drugReservoir
        -actuator
        -microneedleArray
        +administerDrug()
    }
    ProcessingUnit "1" -- "1" MicroneedlePatch : sends trigger signal

2. Operational Parameter Expansion

2.1. High-Pressure Environment Closed-Loop Decompression Sickness Mitigation

Enabling Description: The system is adapted for deep-sea divers to monitor cardiac stress and mitigate decompression sickness (DCS). The wearable device is housed in a pressure-resistant titanium casing rated to 30 ATM. The processor analyzes heart sounds for specific acoustic signatures associated with the formation of nitrogen microbubbles in the bloodstream, which can cause cardiac strain and DCS. The extracted feature is not just PSD but also higher-order spectral analysis (bispectrum) to detect non-linearities caused by bubble transit. The reservoir contains a PFO (Perfluorooctane) emulsion, a high-efficiency oxygen carrier. If acoustic markers of high bubble load are detected, the system administers a micro-dose of the PFO emulsion via an integrated intramuscular autoinjector. This enhances off-gassing of nitrogen, providing a real-time, closed-loop intervention to reduce DCS risk during ascent.

Diagram:

flowchart TD
    A[Acoustic Sensor in Pressure Housing] --> B{Processor};
    B -- Bispectrum Analysis --> C[Bubble Signature Detection];
    C -- Threshold Exceeded --> D{Trigger Autoinjector};
    D --> E[Administer PFO Emulsion];
    subgraph Diver-Worn System
        A --> B --> C --> D --> E
    end

3. Integration with Emerging Tech

3.1. AI-Driven Reinforcement Learning for Therapy Optimization

Enabling Description: The drug administration logic is controlled by a reinforcement learning (RL) agent running on the device's processor. The "state" is the vector of acoustic features extracted from the heart sound. The "action" is the dosage and timing of the compound administered from the reservoir. The "reward" is a function that positively scores the return of the acoustic features to a predetermined healthy baseline and negatively scores the amount of drug used (to prevent overuse). The RL agent, using an algorithm like Q-learning or Deep Q-Networks (DQN), learns a personalized policy for the patient over time, optimizing the therapy to be maximally effective with minimal dosage by observing the acoustic response to each administration event.

Diagram:

graph LR
    State[Acoustic Feature Vector] --> Agent[RL Agent];
    Agent -- Action --> Env[Patient & Treatment System];
    Env -- Reward --> Agent;
    Env -- Next_State --> State;
    subgraph Legend
        direction LR
        State_Def[State: Current heart sound features]
        Action_Def[Action: Drug dosage/timing]
        Reward_Def[Reward: Return to baseline vs. drug cost]
    end

3.2. Blockchain for Verifiable Treatment Audit Trail

Enabling Description: To ensure regulatory compliance and patient safety for the automated treatment system, every critical event is logged as a transaction on a private, permissioned blockchain (e.g., Hyperledger Fabric). When the processor measures a heart sound, a hash of the raw data and the extracted feature vector is created and stored. When a treatment is administered, a new transaction is created containing a timestamp, the feature vector that triggered the event, the dosage administered, and a cryptographic link to the previous transaction. This creates an immutable, tamper-proof, and auditable log of the device's diagnostic and therapeutic actions, which can be securely reviewed by clinicians and regulatory bodies without compromising patient privacy.

Diagram:

erDiagram
    BLOCK {
        string BlockHash PK
        string PreviousBlockHash
        string Timestamp
        string MerkleRoot
    }
    TRANSACTION {
        string TxID PK
        string Timestamp
        string FeatureVectorHash
        string Dosage
        string TriggerCondition
    }
    BLOCK ||--o{ TRANSACTION : contains

4. The "Inverse" or Failure Mode

4.1. Fail-Safe Administration via "Heartbeat" Signal

Enabling Description: The drug administration mechanism is designed to be inherently fail-safe. The main processor, in addition to its analytical tasks, generates a continuous, encrypted "heartbeat" signal (e.g., a 1 Hz square wave with a rolling code) as long as it is operating correctly. This signal is sent to a separate, low-power safety microcontroller that directly controls the gate of the administration device (e.g., the power to the pump or microneedle actuator). If this heartbeat signal ceases for any reason—such as the main processor crashing, the device detaching from the skin (leading to a poor signal), or a battery failure—the safety microcontroller immediately and irrevocably shuts down the administration pathway. This prevents any possibility of accidental overdose due to software or hardware failure.

Diagram:

sequenceDiagram
    participant MainPU as Main Processor
    participant SafetyMCU as Safety Microcontroller
    participant Actuator as Drug Delivery Actuator

    loop Normal Operation
        MainPU->>SafetyMCU: Encrypted Heartbeat Signal
        SafetyMCU->>Actuator: Maintain 'Enable' State
    end

    break Main Processor Fails
        MainPU-xSafetyMCU: Heartbeat Signal Stops
        SafetyMCU->>SafetyMCU: Timeout Exceeded
        SafetyMCU-xActuator: De-assert 'Enable' Signal (Fail-Safe)
    end

Combination Prior Art with Open-Source Standards

Integration with TensorFlow Lite for Microcontrollers: The entire signal processing pipeline of US 11,484,284—specifically the sum of squares error segmentation, Fast Fourier Transform for PSD calculation, and k-means clustering algorithm—is implemented using the open-source TensorFlow Lite for Microcontrollers library. The models are quantized to 8-bit integers for execution on a low-power ARM Cortex-M4 microcontroller. This enables the complete system of claim 17 to operate on a coin-cell battery for weeks, making the device smaller and more accessible. The implementation details, including the C++ code for feature extraction and model inference, are hereby disclosed.
Integration with HL7 FHIR Standard: The output of the system is formatted according to the Health Level Seven (HL7) Fast Healthcare Interoperability Resources (FHIR) R4 standard. The calculated Power Spectral Density is stored as a DiagnosticReport resource with a series of Observation components representing the power at different frequency bands. The output of the unsupervised machine learning model (the disease stage) is an Observation resource with a LOINC code. The recommendation for therapy (Claim 1) or the record of an automated administration (Claim 17) is captured as a MedicationRequest or MedicationAdministration resource, respectively. This enables plug-and-play interoperability with any compliant Electronic Health Record (EHR) system.
Use of MQTT Protocol for Data Telemetry: For applications involving remote monitoring, the wearable device acts as an MQTT client. After local processing, it publishes a JSON payload containing key extracted features (e.g., S1/S2 duration, PSD AUC, dominant frequency, assigned cluster ID) to a specific topic on an MQTT broker. This lightweight, open-source publish/subscribe protocol minimizes bandwidth and battery consumption. A secure connection is established using TLS, and the payload is structured for easy ingestion by cloud-based analytics platforms for longitudinal patient tracking or population health studies. Raw audio data is not transmitted to preserve privacy.