⚡ electrical engineering

Towards Non-Invasive Multiparameter Monitoring from a Single Fingertip Photoplethysmogram: Prototype Design and Preliminary Evaluation of Heart Rate, SpO2, Blood Pressure, and Glucose Estimation

This paper presents the design and preliminary evaluation of a low-cost, single-fingertip photoplethysmogram prototype capable of estimating heart rate, SpO2, blood pressure, and blood glucose, demonstrating feasibility through a small-cohort study where Random Forest regression achieved the most accurate glucose predictions, while emphasizing the need for larger, subject-independent validation before clinical application.

Original authors: Muhammad Noman, Haseeb Ur Rehman, Muhammad Ahsen, Rida Afzal, Muhammad Hamza Zulfiqar, Nida Iqbal

Published 2026-10-20

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Muhammad Noman, Haseeb Ur Rehman, Muhammad Ahsen, Rida Afzal, Muhammad Hamza Zulfiqar, Nida Iqbal

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Millions of people around the world manage chronic conditions like diabetes and high blood pressure by relying on tools that are often painful, expensive, or cumbersome. Checking blood sugar typically requires pricking a finger to draw a drop of blood, while monitoring blood pressure involves wrapping a tight, inflatable cuff around the arm. These methods work, but the discomfort and inconvenience can make people skip their daily checks, leaving dangerous health trends undetected. Scientists have long hoped to replace these invasive tools with a single, comfortable sensor that could read vital signs from a fingertip without breaking the skin. The key to this idea lies in a technique called photoplethysmography, which uses light to measure how blood flows through tiny vessels. By shining light into the skin and measuring how much is absorbed or reflected, devices can detect the rhythmic pulse of blood volume changes. This optical signal already powers the familiar pulse oximeters seen in hospitals and gyms, which tell us heart rate and oxygen levels. The challenge has been to see if this same stream of light data could also reveal blood pressure and glucose levels, the two most critical numbers for managing heart disease and diabetes.

A team of researchers at the University of Engineering and Technology in Lahore set out to build a low-cost prototype to test exactly this possibility. They constructed a small device that clamps onto the tip of a person's finger, housing a sensor that flashes red and infrared light into the skin. This sensor, combined with a simple microcontroller, captures the subtle waves of light that bounce back as blood pulses through the finger. The researchers then fed this raw data into a computer program designed to clean up the signal and extract specific patterns. They tested their system on thirty healthy volunteers who had fasted overnight, comparing the device's estimates against standard medical tools: a finger-prick glucose meter, a traditional blood pressure cuff, and a commercial pulse oximeter. The goal was not to create a finished medical product, but to see if the basic physics and mathematics behind the idea could even work in a real-world setting.

The results for heart rate and oxygen saturation were reassuringly close to the standard devices. The prototype successfully tracked the heartbeat and oxygen levels with a high degree of agreement, confirming that the hardware could capture the essential pulse information accurately. For blood pressure, the team used a different approach, analyzing the timing and shape of the pulse wave rather than just its speed. In a single test case, the device estimated a blood pressure reading that differed from the standard cuff by only about 1.4 millimeters of mercury, a remarkably small gap. However, the researchers were careful to note that this was a single observation on a resting subject, and the method used fixed settings that might not work for everyone without individual adjustment.

The most complex part of the study involved estimating blood glucose. Unlike heart rate or oxygen, there is no direct, simple optical link between sugar levels and the shape of a light pulse. Instead, the researchers hypothesized that changes in blood sugar might subtly alter the blood's thickness or the way red blood cells move, which would in turn change the light signal in detectable ways. To find these subtle clues, they broke the pulse signal down into dozens of different mathematical features, such as the energy of the wave, its rhythm, and its overall shape. They then fed these features into three different types of computer learning models to see which one could best guess the blood sugar level.

Among the models tested, a method known as a Random Forest, which works by building many decision trees and combining their answers, performed the best. On this small group of thirty people, the model's estimates were off by an average of roughly 11 milligrams per deciliter, a significant improvement over the other models it was compared against. The system also managed to correctly categorize the blood sugar levels into low, normal, or high ranges about 74 percent of the time. While these numbers might sound promising, the researchers emphasize that this success is preliminary and limited by the size of the study. The group of volunteers was small, and the data came from a single session, meaning the model might have simply memorized the specific patterns of these thirty people rather than learning a universal rule for blood sugar.

The paper concludes that while the prototype proves the concept is feasible, it is far from ready for clinical use. The researchers explicitly state that their results do not prove the device can measure glucose accurately enough for medical decisions, nor can it yet replace a finger-prick test. The small sample size, the lack of testing on people with diagnosed diabetes or low blood sugar, and the absence of rigorous validation methods mean that the current findings are merely a proof of concept. The device serves as an inexpensive research platform that demonstrates light can carry information about blood sugar, but turning that potential into a reliable medical tool will require much larger studies, testing across diverse populations, and a deeper understanding of how other factors like age and skin tone might confuse the readings. For now, the work stands as a hopeful step toward a future where vital signs can be monitored continuously and painlessly, but the path to a working solution remains long and uncertain.

Technical Summary: Non-Invasive Multiparameter Monitoring from a Single Fingertip PPG

Problem Statement
Cardiovascular disease and diabetes represent significant global health burdens, yet routine monitoring of blood pressure (BP) and blood glucose (BG) relies on methods that hinder patient adherence. Standard cuff-based oscillometry provides intermittent rather than continuous readings and can be uncomfortable, while capillary glucose testing requires painful fingersticks and consumable strips. Although Photoplethysmography (PPG) is established for heart rate (HR) and oxygen saturation (SpO2), its application to cuffless BP and non-invasive glucose estimation remains scientifically unsettled. Challenges include the indirect physical link between glucose and PPG waveforms, the influence of confounding physiological variables (e.g., age, vascular stiffness), and the prevalence of inflated accuracy claims due to small datasets and data leakage.

Methodology
The authors designed and evaluated a low-cost prototype capable of deriving HR, SpO2, systolic/diastolic BP (SBP/DBP), and BG estimates from a single fingertip PPG signal.

Hardware Architecture: The system utilizes a MAX30102 dual-wavelength sensor (660 nm red, 940 nm infrared) integrated with an ATmega328P-based Arduino Nano. The sensor is housed in a 3D-printed, light-shielding enclosure designed to minimize ambient light interference and standardize contact pressure on the right middle fingertip. Data acquisition occurs at 100 samples/second via I²C, with raw data streamed to a PC for offline processing.
Signal Conditioning: The raw PPG signal undergoes baseline removal via lower-envelope detrending (connecting trough points of successive beats) followed by zero-phase 4th-order Butterworth low-pass filtering (8 Hz cutoff) to preserve fiducial points while removing respiratory and high-frequency noise.
Parameter Estimation Algorithms:
- HR & SpO2: HR is calculated from the mean peak-to-peak interval. SpO2 is derived using the ratio of ratios ( $R$ ) of the AC/DC components of the red and infrared channels, applied to a linear approximation formula ( $SpO2 = 110 - 25R$ ).
- Cuffless BP: A proof-of-concept empirical model uses first-derivative timing features (systolic time, diastolic time) and amplitude features. SBP and DBP are mapped using fixed coefficients ( $\alpha, \beta$ ) derived for resting adults, without per-subject calibration.
- Glucose Estimation: A multi-domain feature set was extracted, including autoregressive coefficients (5th order), Teager–Kaiser energy (mean, variance, IQR, skewness), spectral Shannon entropy, log-energy, and single-pulse morphological features (energy, peak amplitude, onset/endset times).
Machine Learning & Evaluation: Three regression models—Linear Regression (LR), K-Nearest Neighbors (KNN, $k=5$ ), and Random Forest (RF, 100 trees)—were trained to predict glucose levels. Performance was evaluated using Root Mean Square Error (RMSE), Mean Absolute Error (MAE), $R^2$ , and three-band classification accuracy (Low: $\le 100$ , Mid: 101–140, High: $> 140$ mg/dL).

Experimental Protocol
The study involved 30 fasting volunteers (aged 20–50) at the University of Engineering and Technology Lahore. A total of 32 recordings (30 seconds each) were collected. Reference measurements were taken simultaneously using a commercial fingertip pulse oximeter, an oscillometric cuff, and an Accu-Chek Active glucometer. The cohort consisted of healthy-to-mixed individuals without independent clinical diagnosis of diabetes; no hypoglycemic subjects were included.

Key Results

HR and SpO2: The prototype demonstrated close agreement with commercial references. In a representative 30-second recording, the estimated HR was ~73 BPM versus 70 BPM on the reference (approx. 4% difference), and SpO2 was estimated at 99% versus the reference value.
Blood Pressure: A single resting trial showed an estimated BP of 118.6/78.6 mmHg compared to 120/80 mmHg from the cuff, a difference of approximately 1.4 mmHg.
Glucose Estimation: The Random Forest (RF) model outperformed LR and KNN.
- RF Performance: RMSE = 11.11 mg/dL, MAE = 7.85 mg/dL, $R^2$ = 0.8655, and 74.19% three-band classification accuracy.
- Comparison: LR yielded an RMSE of 24.31 mg/dL ( $R^2$ = 0.3557), and KNN yielded 26.45 mg/dL ( $R^2$ = 0.2374).
- Limitations in Results: The authors note that the high $R^2$ and low error metrics are likely optimistic due to the small dataset size (32 recordings) and the lack of subject-independent validation (potential data leakage between training and test sets).

Significance and Claims
The paper positions this work strictly as a feasibility study and a preliminary evaluation, explicitly avoiding clinical claims.

Feasibility of Integration: The study demonstrates that a low-cost, single-sensor hardware platform can successfully acquire and condition PPG signals to derive multiple physiological parameters simultaneously.
Feature Utility: The results suggest that a combination of spectral, non-linear (Teager–Kaiser), and morphological features can be associated with reference glucose values in a small, fasting cohort, with Random Forest capturing non-linear interactions better than linear or distance-based models.
Modesty of Claims: The authors emphasize that the current results do not validate the system for clinical use. They highlight that:
- The glucose model may be learning correlates (e.g., vascular tone, age) rather than glucose itself.
- The lack of subject-independent validation (e.g., leave-one-subject-out) renders performance estimates unreliable.
- The narrow physiological range of the cohort (no hypoglycemia, limited hyperglycemia) prevents assessment of clinical extremes.
- Standard clinical metrics (Clarke/Parkes error grids, MARD, ISO 15197 compliance) were not met or fully evaluated.

Conclusion
The prototype serves as an inexpensive research platform for collecting larger datasets. The authors conclude that before any clinical claims can be made, future work must include larger, multi-day studies with diagnosed diabetic and hypoglycemic populations, subject-independent validation protocols, and rigorous error-grid analysis.

Technical Summary: Non-Invasive Multiparameter Monitoring from a Single Fingertip PPG

More like this