FORM-2

THE PATENTS ACT,1970 (39 OF 1970)

&

THE PATENTS RULES, 2003

COMPLETE SPECIFICATION
(SECTION 10, RULE 13)

TITLE

“NON-INVASIVE HEMOGLOBIN SCREENING SYSTEM USING MULTI-

WAVELENGTH VIDEO ANALYSIS AND METHOD THEREOF”

APPLICANT(S)

ARUNITA TUSHAR JAGZAPE, an Indian National of Department of

Physiology, AIIMS Raipur, Great Eastern Rd, opposite Gurudwara,
AIIMS Campus, Tatibandh, Raipur, Chhattisgarh 492099, INDIA;

DR. GAGAN RAJ GUPTA, Indian National of Computer Science and

Engineering, Joint Faculty, Mechatronics, Indian Institute of
Technology, Bhilai, Chhattisgarh 491001, INDIA; &

DR. TUSHAR B JAGZAPE, an Indian National of Department of

Paediatrics, AIIMS Raipur, Great Eastern Rd, opposite Gurudwara,
AIIMS Campus, Tatibandh, Raipur, Chhattisgarh 492099, INDIA

The following specification particularly describes the invention and the

manner in which it is to be performed

1
 FIELD OF INVENTION
The present invention relates to a Non-Invasive Hemoglobin Screening
system and method therefor. More particularly, the present invention is
related to an improved screening system for estimation of haemoglobin. The
5 said non-invasive screening system is also configured for identifying types of
anemia using multi-wavelength video analysis and machine learning.

BACKGROUND ART
In anemia, haemoglobin (Hb) goes below its normal level in RBCs which
10 results decreased oxygen carrying capacity leads to subsequent severe
medical condition.

In existing art, most common invasive methodology for Hb test is based on

use of blood sample drawn through a finger prick or venipuncture blood
15 draw. In the clinical setup, health assistants collect a 3mL sample of blood
for a complete blood count (CBC) test including haemoglobin test. The
system requires withdrawal of blood, either through venous sample or
capillary blood with sterilization issues, also needs various chemical
reagents, beside this, this may be procedural inaccuracies and
20 interpretation difficulties and apprehension and fear of prick among few
enforcing the need for non-invasive devices.

The evolution of non-invasive hemoglobin measurement has seen a

transition from invasive methods to optical-based approaches, and further
25 to sophisticated spectroscopic techniques. However, while these
advancements are promising, challenges remain. High costs, requirements
of chemicals and inconsistent results continue to limit the widespread
adoption of these technologies, particularly for personal use.

Despite these advancements, there remains a need for more affordable,

30 reliable, and user-friendly solutions. Our research aims to build on this
foundation by exploring the potential of portable non-invasive hemoglobin

2
 monitoring; leveraging machine learning and innovative light-based
techniques to address the limitations of existing methods.

Researchers have investigated non-invasive techniques for Hb estimation

using optical instruments such as spectrometer, hyperspectral camera etc.
5 which seems to be good in view of accuracy, but these techniques are highly
expensive and are complex in use and needs a specialist in order to perform
the system. US5377674 discloses a non-invasive in-vitro method for
measurement of haemoglobin but this system is complex in nature and is
associated with issue of accuracy. Further portability is another issue for
10 this kind of system. Some researchers have also investigated the use of
estimating Hb levels using images captured by smartphones under
illumination from LED lights of various frequencies. Wang et al. developed a
system called HemaApp [Wang EJ, Li W, Hawkins D, Gernsheimer T, Norby-
Slycord C, Patel SN. HemaApp: Noninvasive blood screening of hemoglobin
15 using smartphone cameras. GetMobile: Mobile Computing and
Communication. 2017; 21(2): 26-30] that predicts blood haemoglobin
concentration using different lighting sources based on an analysis of the
fingertip video data captured using a smartphone. Their setup is complex in
nature and has dependencies on availability of a smartphone. In another
20 study, fingertip videos captured by a smartphone while turning ON the flash
was used to estimate Hb concentration [Hasan MK, Haque M, Sakib N, Love
R, Ahamad SI. Smartphone-based Human Hemoglobin Level measurement
analyzing pixel intensity of a fingertip video on different colour spaces. Smart
Health. 2018; 5-6: 26-39.]. Most of these studies concluded the result based
25 on the very small sample size (<50 patients). Yet other techniques are failed
to identify the type of anemia which is another important aspect in this field.

Therefore, there is a need to provide an improved system which can solve

the aforesaid deficiencies of the existing art.

Hence, the gap lies first in the: a) testing component that will provide a true
30 picture of prevalence of anaemia that will further enhance the treatment
component b) availability of portable, easy to comprehend, non-invasive
devices that are reagent free and fairly accurate in estimation of
3
 haemoglobin and RBC to detect whether anaemia is present and also guide
towards the type of nutritional anaemia that is mostly the cause.

Therefore, an in-vitro non-invasive Class A, portable, reliable, valid, easy to

perform, cost effective, reagent free, operable by all, pre-analytical errorless
5 Hb estimation device is proposed which would help in diagnosis of anaemia
and type (IDA (iron deficiency anaemia) or Megaloblastic anaemia that would
further lead to its timely treatment especially taking into consideration the
rural population end users.

10 SUMMARY OF INVENTION

Hemoglobin is an essential protein found in red blood cells (RBCs),

responsible for transporting oxygen to all parts of our body. Types of anemia
like Iron-deficiency anemia are very common in women. Iron-deficiency
anemia is more prevalent globally along with vitamin-deficiency anemia
15 caused by low B12. Hemoglobin levels are measured by pricking and blood
extraction at laboratories by traditional Sahli’s method where, HCl solution
when added to blood sample and kept for ~10 minutes creates a solution of
acid hematin which is then diluted and the colour is matched with the
coloured glass rods of the comparator. The intensity of colour of this
20 solution is directly proportional to the hemoglobin level of the individual.
With the advancements in Image Processing and Machine Learning, we
propose a method to replace the traditional method by correlating pixel
intensities of the captured video of patients finger in multiple wavelengths of
light including Infrared.

25 Therefore, the principle object of the present invention is to provide a

compact system that can estimate the level of haemoglobin accurately. The
instant system consists of two compartments being connected to each other,
finger holder, microcontroller board, power supply plug, visual camera and
with IR camera having plurality of wavelength.

30 In an embodiment, the object (patient) is requested to put his right-hand

index-finger in the holder in order to get a video of the same which is further

4
 subjected for different wavelength (580-650). In an embodiment, the
duration of such video is 30 seconds.

In an embodiment, the aforesaid mechanism is based on the absorption of

specific light which are placed separately, the red wavelength provides the
5 level of haemoglobin with the help of the captured video. The 30 seconds
video, undergoes video processing via python opencv libraries, in which one
can analyze each frame in the video and note the Red, Blue, Green, Hue,
Saturation, Value, Light, A channel and Grayscale values of each pixel in
the given Region of Interest in every frame. Each frame has a histogram,
10 depicting the frequency of each intensity value in a particular frame. The
multiple frames data have different values of intensity of each colour space
of the same pixel, therefore an average histogram is built for all the frames.
The average histogram depicts the average occurrence of each intensity
value throughout the video. Thus, we quantify the video in form of nine
15 features Red, Blue, Green, Hue, Saturation, Value, Light, A channel and
Grayscale (RGBHCVLAGray). These 9 average histograms of
RGBHCVLAGray undergo pre-processing and we have the 9 features used as
an input to our Neural Network. The Deep Learning model learns the
relationship between said 9 features and Hb count, finally after training
20 efficiently we can predict the Hb count on unseen data (video). In order to
make it compact and less space occupying, Arduino NANO 33 BLE Sense
microcontroller is used which is one fifth the size of current board, which
will be the microcontroller too.

Therefore such as herein described there is provided a non-invasive

25 hemoglobin screening device comprising of a housing with two connected
compartments and a finger holder configured to receive a patient’s index
finger; a visual camera and an infrared camera configured to capture a 40-
second video of the finger under multiple wavelengths including 570 nm,
620 nm, 720 nm, and infrared; a microcontroller configured to process the
30 video and extract pixel intensity values in nine color spaces including Red,
Blue, Green, Hue, Saturation, Value, Light, A channel, and Grayscale; a
neural network model executed by the microcontroller, the model trained to

5
 predict hemoglobin levels and classify anemia types based on a feature
vector derived from average histograms of the nine color spaces; and a
power supply configured to support continuous operation.

5 In another aspect herein disclosed is a method for non-invasive hemoglobin

screening, comprising the steps of capturing a 40-second video of a patient’s
index finger using a visual camera and an infrared camera under
wavelengths of 570 nm, 620 nm, 720 nm, and infrared; processing the video
to extract pixel intensity values in nine color spaces including Red, Blue,
10 Green, Hue, Saturation, Value, Light, A channel, and Grayscale; generating
average histograms for each color space and converting them into a 2304-
element feature vector; applying a neural network model to the feature
vector to predict hemoglobin levels and classify anemia types; and
outputting the hemoglobin level and anemia classification.
15
In yet another aspect herein disclosed is a method for processing video data
for hemoglobin screening, comprising the steps of receiving a video of a
patient’s finger captured under multiple wavelengths; extracting pixel
intensity values in at least five color spaces from a region of interest in each
20 frame; generating a feature matrix from histograms of the intensity values;
flattening the feature matrix into a vector; and applying a trained neural
network to the vector to output a hemoglobin level and an anemia
classification.

25 BRIEF DESCRIPTION OF THE ACCOMPANYING DRAWINGS:

Fig. 1 illustrates working of the system in accordance with the present
invention;
Fig. 2 illustrates 3-d structure of the system and finger-hole placement in
accordance with the present invention;
30 Fig. 3 (a), 3(b), 3(c) and 3(d) illustrates the top view, front view (where the
finger is placed), back view of the base and top view of the system
respectively in accordance with the present invention;

6
 Fig. 4(a) illustrates the correlated pixel intensity representation in nine
different color spaces, in accordance with the present invention;
Fig. 4(b) illustrates the device used for collection of Data in accordance with
the present invention;
5 Fig. 5 illustrates the image of conversion of the feature matrix gained from
the histogram to the feature vector in accordance with the present invention;
Fig. 6 illustrates an image of the circuit configured for said device in
accordance with the present invention;
Fig. 7 illustrates an Hb level prediction in accordance with the present
10 invention.

DETAILED DESCRIPTION:

The present invention provides a non-invasive hemoglobin screening system

as shown in Fig. 1, including a device that correlates pixel intensity values
in multiple color spaces with hemoglobin levels and anemia types. The
15 device leverages multi-wavelength video analysis and machine learning to
achieve high accuracy, portability, and ease of use, addressing the
limitations of existing invasive and non-invasive methods. The following
sections describe the device’s structure, operation, data processing, machine
learning model, and embodiments in exhaustive detail, with references to
20 the accompanying drawings.

Device Structure

As illustrated in Figure 2, the device comprises a compact, portable housing

with two connected compartments, designed to minimize size and weight for
use in rural and resource-limited settings. The housing is constructed from
25 biocompatible, medical-grade polycarbonate, ensuring durability and safety
per Class A medical device standards. The compartments are seamlessly
integrated to protect internal components from environmental factors such
as dust and moisture (IP54 rating).

Finger Holder

7
 The finger holder (Figure 2) is a cylindrical slot (diameter: 20 mm, depth: 30
mm) located on the top surface of the device as shown in Fig 3 (a) – (d),
designed to securely position the patient’s right index finger during video
capture. The holder is lined with a soft, hypoallergenic silicone pad (2 mm
5 thickness) to enhance comfort and prevent slippage. An adjustable tension
mechanism ensures a snug fit for varying finger sizes (15–25 mm diameter).
The holder’s interior is optically sealed to prevent external light interference,
ensuring accurate video capture.

Microcontroller

10 As shown in Fig 6, the said system is powered by a Raspberry Pi

microcomputer, offering significantly greater computational capabilities
compared to standard microcontroller boards. The Raspberry Pi features a
quad-core ARM Cortex-A72 processor (1.5 GHz), up to 4 GB of RAM, and
supports both onboard Wi-Fi and Bluetooth for wireless data transmission
15 to external devices (e.g., smartphones or cloud servers). This makes it
suitable for real-time video processing and machine learning inference.

Power Supply

A B0198 Arducam Stereo power supply provides stable 5V DC power via a

USB-C connector, supporting continuous operation for up to 12 hours on a
20 2000 mAh lithium-ion battery. The power supply includes a wireless
charging coil (Qi standard) for user convenience. A power management unit
(PMU) regulates voltage and protects against overcharging, ensuring
compliance with IEC 60601-1 medical device safety standards.

Cameras

25 The device integrates two cameras for multi-wavelength video capture:

Visual Camera: A B0198 Arducam Stereo USB Camera with a 2 MP sensor

captures video under visible light at 570 nm, 620 nm, and 720 nm. The
camera features a 60° field of view, f/2.8 aperture, and adjustable focus (2–

8
 10 cm). It is equipped with a triple-wavelength LED array (570 nm ± 5 nm,
620 nm ± 5 nm, 720 nm ± 5 nm) for precise illumination.

Infrared Camera: An OV2710 OmniVision CMOS sensor captures 10-second

IR video (850 nm ± 10 nm). The IR camera has a 1 MP resolution, 70° field of
5 view, and f/2.0 aperture, optimized for low-light conditions.

Both cameras are mounted within the first compartment, aligned with the
finger holder’s optical path (Figure 3(b)). A dichroic filter ensures
wavelength-specific light capture, minimizing cross-talk.

Housing

10 The housing (Figures 3(a)-3(d)) measures 100 mm × 60 mm × 30 mm and

weighs 150 g, ensuring portability. It includes a 2.8-inch OLED touchscreen
display (resolution: 240 × 320 pixels) for real-time results and user
interaction. Ventilation slots and a thermal management system (copper
heat sink) prevent overheating during continuous operation. The housing is
15 ergonomically designed with rounded edges and a non-slip grip for handheld
use.

Operation

As depicted in Figure 4(a), the device operates through a sequence of steps

to capture and process video data for hemoglobin estimation and anemia
20 classification:

Preparation: The patient sanitizes their right index finger to remove

contaminants (e.g., dirt, oil) that could affect optical measurements. The
finger is placed in the holder, and the device is activated via the touchscreen
or a physical button (Figure 3).

25 Video Capture: The device records a 40-second video at 20 frames per

second (fps), totaling 800 frames:

9
 First 30 seconds (600 frames): The visual camera captures video under
sequential illumination at 570 nm (10 seconds), 620 nm (10 seconds), and
720 nm (10 seconds).

Last 10 seconds (200 frames): The IR camera captures video under 850 nm
5 illuminations. The cameras are synchronized to ensure consistent frame
rates and exposure times (1/60 s per frame).

Data Transmission: The video is stored in onboard flash memory (16 GB)
and optionally transmitted via BLE to a paired device for further analysis or
cloud storage. Data encryption (AES-256) ensures compliance with HIPAA
10 and GDPR standards.

Data Processing

The captured video is processed using Python OpenCV libraries (version

4.5.5) on the microcontroller or an external computing device for enhanced
performance. The processing pipeline, illustrated in Figure 4(a), includes the
15 following steps:

Region of Interest (ROI) Selection

To minimize noise from edge effects (e.g., light leakage), a 360-pixel ROI (20
× 18 pixels) is defined in the center of each frame. The ROI corresponds to
the finger’s ventral surface, rich in capillary blood vessels, ensuring optimal
20 hemoglobin-related optical signals. The ROI is dynamically adjusted (±5
pixels) to account for slight finger movements.

Frame Analysis

Each of the 800 frames is analyzed to extract pixel intensity values in nine
color spaces: Red, Blue, Green, Hue, Saturation, Value, Light, A channel
25 (from Lab color space), and Grayscale. The intensity values (0–255) are
stored in CSV files for each frame. For example:

10
 Red Channel: Reflects hemoglobin’s absorption at 570–620 nm, critical for
concentration estimation.

IR Channel: Penetrates deeper tissues, enhancing accuracy for anemia.

Hue and Saturation: Capture color variations due to blood oxygenation.

5 A histogram (256 bins) is generated for each color space per frame,
representing the frequency of intensity values (Figure 4(a)). The histogram is
normalized to account for variations in illumination intensity.

Average Histogram

The histograms from all 800 frames are averaged to create a single
10 histogram per color space, summarizing the distribution of intensity values
across the video. For example, the Red channel’s average histogram
indicates the mean frequency of each intensity value (0–255) observed in the
ROI over 40 seconds. This reduces noise and captures temporal stability in
optical signals.

15 Data Augmentation

To enhance model robustness, the ROI is augmented by:

Rotation: Rotating the ROI by 1° to 20° in 1° increments, generating

20 representations per video.

Size Variation: Analyzing ROIs of 300, 360, and 400 pixels to capture
20 spatial variability.

Temporal Subsampling: Extracting subsets of frames (e.g., every 5th

frame) to simulate variable video lengths.

This augmentation increases the dataset from 100 patient videos to 2000
data points, improving generalization.

25 Feature Extraction

11
 The nine average histograms form a 9 × 256 matrix, where each column
represents a color space (R, G, B, H, S, V, L, A, Grayscale) and each row
corresponds to an intensity bin (0–255). The matrix is flattened into a 1 ×
2304 feature vector (Figure 5), which encapsulates the video’s optical
5 characteristics. The vector is normalized (mean = 0, standard deviation = 1)
to ensure consistent input to the neural network.

Machine Learning Model

The device employs a four-layer feed forward neural network, implemented

using TensorFlow (version 2.8.0), to predict hemoglobin levels and classify
10 anemia types. The model’s architecture and training are detailed below:4.1

Model Architecture

• Input Layer: Accepts the 2304-element feature vector.

• Hidden Layers:

o Layer 1: 512 neurons, ReLU activation, dropout rate 0.3.

15 o Layer 2: 256 neurons, ReLU activation, dropout rate 0.2.

o Layer 3: 128 neurons, ReLU activation.

Output Layer:

o For regression (hemoglobin prediction): 1 neuron, linear activation.

o For classification (anemia typing): 3 neurons (non-anemic, microcytic,

20 macrocytic), softmax activation.

Training Parameters

• Optimizer: Adam (learning rate: 0.001, beta1: 0.9, beta2: 0.999).

• Loss Function:

12
 o Regression: Mean squared error (MSE).

o Classification: Categorical cross-entropy.

• Batch Size: 32.

• Epochs: 30.

5 • Regularization: L2 regularization (lambda = 0.01) to prevent

overfitting.

Dataset

The model is trained on 2000 data points derived from 100 patient videos,
augmented as described in Section 3.4. The dataset is split as follows:

10 • Training Set: 1400 samples (70%).

• Validation Set: 400 samples (20%).

• Test Set: 200 samples (10%), kept unseen during training.

Each sample includes the 2304-element feature vector and ground-truth

hemoglobin levels (obtained via laboratory tests, e.g., CBC). Anemia labels
15 (non-anemic, microcytic, macrocytic) are assigned based on clinical
diagnoses.

Performance

After 30 epochs, the model achieves:

Regression: Mean square error of 1.6055 on the test set, corresponding to a

20 mean absolute error of ±0.87 g/dL. For a patient with a laboratory-tested
hemoglobin level of 14 g/dL, the model predicts 14 ± 0.87 g/dL (Figure 7).

Classification: Accuracy of 92% on the test set, with performance metrics

reported via confusion matrices and ROC curves. The model distinguishes
microcytic (e.g., IDA) and macrocytic (e.g., megaloblastic) anemia based on

13
 spectral absorption patterns, particularly in the red (620 nm) and IR (850
nm) wavelengths.

Inference

During operation, the trained model is deployed on the Arduino Nano’s flash
5 memory or an external device (e.g., Raspberry Pi 4) for real-time inference.
The feature vector is processed in <1 second, and results are displayed on
the OLED screen or transmitted via BLE.

Anemia Type Identification

The model identifies anemia types by analyzing spectral absorption patterns:

10 • Microcytic Anemia: Characterized by smaller RBCs, showing higher

absorption at 570–620 nm due to denser hemoglobin packing.

• Macrocytic Anemia: Characterized by larger RBCs, showing distinct IR

absorption (850 nm) due to altered cell morphology.

• Non-Anemic: Balanced absorption across wavelengths, with no

15 extreme deviations.

The IR camera’s data enhances accuracy by capturing subsurface blood

characteristics, complementing visible light data.

Circuit Configuration

As shown in Figure 6, the circuit integrates:

20 • Microcontroller: Raspberry Pi 4B, connected via I2C and SPI

interfaces.

• Cameras: Visual and IR cameras, powered by the B0198 Arducam

Stereo supply (5V, 500 mA).

14
 • LED Array: Triple-wavelength LEDs (570 nm, 620 nm, 720 nm) and IR
LED (850 nm), controlled by PWM signals.

• Display: OLED screen, interfaced via I2C.

• Battery: Lithium-ion battery with PMU, connected to a USB-C port.

5 The circuit is optimized for low power consumption (average: 2.5 W), with a
sleep mode reducing consumption to 50 mW when idle.

Advantages

• Accuracy: ±0.87 g/dL for hemoglobin estimation, surpassing many

non-invasive methods.

10 • Anemia Typing: Identifies microcytic and macrocytic anemia,

enhancing diagnostic utility.

• Portability: Compact (100 mm × 60 mm × 30 mm, 150 g), ideal for

rural healthcare.

• Reagent-Free: Eliminates chemical dependencies, reducing costs and

15 complexity.

• Ease of Use: Operable by non-specialists with minimal training (<5

minutes).

• Compliance: Meets Class A medical device standards (IEC 60601-1)

and data privacy regulations (HIPAA, GDPR).

20 Embodiments

Handheld Device:

o Includes a touchscreen display and onboard processing for standalone

operation.

o Suitable for clinics and mobile health units.

15
 Cloud-Connected System:

o Transmits raw video data to a cloud server for advanced processing

(e.g., ensemble models).

o Enables telemedicine and population-level anemia screening.

5 Pediatric Variant:

o Smaller finger holder (10–15 mm diameter) for children.

o Adjusted ROI (200 pixels) and model retrained on pediatric datasets.

5. Multi-User Device:

o Includes a fingerprint sensor for patient identification.

10 o Stores up to 100 patient profiles for longitudinal monitoring.

Example

A healthcare worker in a rural clinic activates the device and instructs a

patient to sanitize their right index finger. The finger is placed in the holder,
and the device captures a 40-second video. The OLED screen displays
15 “Processing” while the microcontroller extracts the feature vector and runs
the neural network. Within 5 seconds, the screen shows:

• Hemoglobin Level: 12.3 ± 0.87 g/dL.

• Anemia Status: Microcytic (IDA suspected). The results are

transmitted to a cloud server, where a physician reviews them and
20 prescribes iron supplements. The process is repeated for 20 patients in a
single session, with the battery lasting the entire day.

Modifications

The disclosure allows for modifications in:

16
 • Materials: Use of ABS plastic or aluminum for the housing.

• Cameras: Higher-resolution sensors (e.g., 5 MP) for enhanced ROI

analysis.

• Wavelengths: Additional wavelengths (e.g., 532 nm) for improved

5 spectral resolution.

• Model: Alternative architectures (e.g., convolutional neural networks)

for video processing. Such modifications remain within the invention’s
scope, provided they achieve non-invasive hemoglobin screening via multi-
wavelength video analysis.

10 Inventive step: The present system provides accurate result with regard to
the level of haemoglobin in a simple manner and also identifies the other
type of anemia for example microcytic and macrocytic anaemia.

The disclosure herein explicitly states that there can be slight change in the
design and configuration in actual to conceive the proposed solution. The
15 foregoing description is a specific embodiment of the present disclosure. It
should be appreciated that this embodiment is described for purpose of
illustration only, and that those skilled in the art may practice numerous
alterations and modifications without departing from the spirit and scope of
the invention. It is intended that all such modifications and alterations be
20 included insofar as they come within the scope of the invention as claimed
or the equivalents thereof.

17
We Claim:-

1. A non-invasive hemoglobin screening device comprising:

a housing with two connected compartments and a finger holder

configured to receive a patient’s index finger;

a visual camera and an infrared camera configured to capture a 40-

second video of the finger under multiple wavelengths including 570 nm,
620 nm, 720 nm, and infrared;

a microcontroller configured to process the video and extract pixel

intensity values in nine color spaces including Red, Blue, Green, Hue,
Saturation, Value, Light, A channel, and Grayscale;

a neural network model executed by the microcontroller, the model

trained to predict hemoglobin levels and classify anemia types based on a
feature vector derived from average histograms of the nine color spaces;
and

a power supply configured to support continuous operation.

2. The non-invasive hemoglobin screening device as claimed in claim 1,

wherein the microcontroller is Raspberry Pi 4B, and the power supply
is a B0198 Arducam Stereo power supply.
3. The non-invasive hemoglobin screening device as claimed in claim 1,
wherein the cameras include a B0198 Arducam Stereo USB Camera
and an OV2710 OmniVision CMOS sensor.
4. The non-invasive hemoglobin screening device as claimed in claim 1,
wherein the video comprises 30 seconds under visible light and 10
seconds under infrared light, captured at 20 frames per second.
5. The non-invasive hemoglobin screening device as claimed in claim 1,
wherein the neural network model achieves a hemoglobin prediction
accuracy of ±0.87 g/dL.

18
6. The non-invasive hemoglobin screening device as claimed in claim 1,
wherein the neural network model classifies anemia as microcytic or
macrocytic based on spectral absorption patterns.
7. The non-invasive hemoglobin screening device as claimed in claim 1,
wherein the housing is constructed from medical-grade polycarbonate
and includes a 2.8-inch OLED touchscreen display.
8. The non-invasive hemoglobin screening device as claimed in claim 1,
wherein the finger holder includes a silicone-lined slot with an
adjustable tension mechanism for varying finger sizes.
9. A system for non-invasive hemoglobin screening, comprising:

a portable device including a finger holder, a visual camera, an

infrared camera, a microcontroller, and a power supply;

a video capture module configured to record a 40-second video of a

patient’s index finger under wavelengths of 570 nm, 620 nm, 720 nm,
and infrared;

a video processing module configured to extract pixel intensity values

in nine color spaces and generate a 2304-element feature vector from
average histograms; and

a deep learning module configured to predict hemoglobin levels and

classify anemia types using the feature vector.

10. The system for non-invasive hemoglobin screening as claimed in

claim 9, further comprising a data augmentation module configured to
rotate a region of interest by 1 to 20 degrees to generate multiple video
representations.
11. The system for non-invasive hemoglobin screening as claimed in
claim 9, wherein the video processing module uses Python OpenCV
libraries to analyze a 360-pixel region of interest in each frame.
12. The system for non-invasive hemoglobin screening as claimed in
claim 9, further comprising a wireless communication module

19
 configured to transmit results to a cloud server via Bluetooth Low
Energy.
13. A method for non-invasive hemoglobin screening, comprising
the steps of:

capturing a 40-second video of a patient’s index finger using a visual

camera and an infrared camera under wavelengths of 570 nm, 620
nm, 720 nm, and infrared;

processing the video to extract pixel intensity values in nine color

spaces including Red, Blue, Green, Hue, Saturation, Value, Light, A
channel, and Grayscale;

generating average histograms for each color space and converting

them into a 2304-element feature vector;

applying a neural network model to the feature vector to predict

hemoglobin levels and classify anemia types; and

outputting the hemoglobin level and anemia classification.

14. The method for non-invasive hemoglobin screening device as

claimed in claim 13, wherein the neural network model is trained on a
dataset of 2000 augmented data points derived from 100 patient
videos.
15. The method for non-invasive hemoglobin screening device as
claimed in claim 13, wherein the video is captured at 20 frames per
second, yielding 800 frames for analysis.
16. The for non-invasive hemoglobin screening device as claimed in
claim 13, further comprising sanitizing the patient’s finger prior to
video capture to remove contaminants.
17. A non-invasive hemoglobin screening device comprising:

a portable housing with a finger holder configured to position a

patient’s finger;

20
 a camera system configured to capture a video of the finger under at
least three visible wavelengths and one infrared wavelength;

a processing unit configured to extract pixel intensity features from

the video in multiple color spaces and generate a feature vector; and

a machine learning model configured to predict hemoglobin levels with

an accuracy of ±0.87 g/dL and classify anemia types based on the
feature vector.

18. The non-invasive hemoglobin screening device as claimed in

claim 17, wherein the camera system includes a triple-wavelength
LED array emitting light at 570 nm, 620 nm, and 720 nm.
19. The non-invasive hemoglobin screening device as claimed in
claim 17, wherein the processing unit normalizes the feature vector to
a mean of zero and a standard deviation of one.
20. The non-invasive hemoglobin screening device as claimed in
claim 17, further comprising a thermal management system including
a copper heat sink to prevent overheating.
21. A method for processing video data for hemoglobin screening,
comprising the steps of :

receiving a video of a patient’s finger captured under multiple

wavelengths;

extracting pixel intensity values in at least five color spaces from a

region of interest in each frame;

generating a feature matrix from histograms of the intensity values;

flattening the feature matrix into a vector; and

applying a trained neural network to the vector to output a

hemoglobin level and an anemia classification.

21
 22. The method for processing video data for hemoglobin screening
as claimed in claim 21, wherein the neural network includes four
layers with ReLU activation functions and is trained using an Adam
optimizer.
23. The method for processing video data for hemoglobin screening
as claimed in claim 21, wherein the feature matrix comprises nine
color spaces and 256 intensity bins per space.

Dated this 23rd day of April, 2025

Arghya Ashis Roy

Patent Agent (IN/PA 2346)
Of Lex-Regia
For the Applicant(s)

To,
The Controller of Patents,
The Patent Office, MUMBAI

22
 ABSTRACT
“NON-INVASIVE HEMOGLOBIN SCREENING SYSTEM USING MULTI-
WAVELENGTH VIDEO ANALYSIS AND METHOD THEREOF”

A non-invasive hemoglobin screening device comprising of a housing with

two connected compartments and a finger holder configured to receive a
patient’s index finger; a visual camera and an infrared camera configured to
capture a 40-second video of the finger under multiple wavelengths
including 570 nm, 620 nm, 720 nm, and infrared; a microcontroller
configured to process the video and extract pixel intensity values in nine
color spaces including Red, Blue, Green, Hue, Saturation, Value, Light, A
channel, and Grayscale; a neural network model executed by the
microcontroller, the model trained to predict hemoglobin levels and classify
anemia types based on a feature vector derived from average histograms of
the nine color spaces; and a power supply configured to support continuous
operation. Fig. 1

23