1 of 10

IECS 2026 · STUDENT RESEARCH PRESENTATION

Predicting Cardiovascular Fitness using

Fingerprint biometrics

A Non-Invasive Hybrid CNN + MLP Framework for Non-Exercise VO2 Max Estimation

Phani Krishna Bulasara · Vuyyuru Hema Sri Kanaka Durga · Chinta Prudhvi Sai Raj · Pilli Rupesh

Department of Artificial Intelligence and Data Science

LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING, ANDHRA PRADESH, INDIA

2 of 10

INTRODUCTION

What’s this VO2 Max is in general ?

Cardiovascular disease remains the leading cause of death worldwide, and maximal oxygen consumption (VO₂ Max) is the gold-standard marker of cardiorespiratory fitness and long-term heart health.
Yet conventional exercise-based VO₂ Max testing is invasive, expensive, and confined to clinical or research settings — out of reach for most people.
Fingerprint ridge patterns encode vascular and epidermal traits that may reflect underlying cardiovascular function — an accessible, non-invasive signal worth exploring.
This work fuses fingerprint images with easily obtained metadata (age, gender, height, weight, BMI, activity level) to estimate VO₂ Max without any lab exercise test.

THE CORE IDEA

Fingerprint Image + Metadata

ResNet-50 (CNN) extracts ridge features → MLP encodes physiological metadata → fused representation predicts VO₂ Max and classifies fitness level.

89.58%

Best-model classification accuracy for fitness-level stratification

IECS 2026 · LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING, ANDHRA PRADESH, INDIA

3 of 10

RELATED WORK

Where the Literature Falls Short

Year / Ref	Data & Approach	Research Gap
2022 [4]	Fingerprint images (~1000 samples) with ResNet-50 / VGG-16	Predicted age only — ignored fitness outcomes
2023 [5]	Anthropometric data (age, height, weight, BMI) with regression	No imaging or biometric modality used
2023 [6]	Wrist-vein biometrics with Siamese / ResNet CNNs	Built for identity recognition, not VO₂ Max
2024 [8]	Smart-healthcare data fusion with deep learning	Lacks a concrete multimodal fusion recipe
2025 [9]	Wearable photoplethysmography with Random Forest	Needs a wrist device — no fingerprint biometrics

Across the field: strong imaging OR strong metadata models exist — but not a fused, label-efficient pipeline for VO₂ Max.

IECS 2026 · LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING, ANDHRA PRADESH, INDIA

4 of 10

PROPOSED SYSTEM

A Multimodal Pipeline, End to End

Construct a dual-modality dataset: fingerprint ridge images + physiologically realistic metadata.
Preprocess images and tabular features with separate, purpose-built pipelines.
Learn jointly with a hybrid CNN + MLP model, fused into shared latent features.
Optimize a weighted multitask loss for simultaneous regression and classification.

Fig. 1 — Hybrid CNN/MLP architecture for cardiovascular-fitness prediction

IECS 2026 · LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING, ANDHRA PRADESH, INDIA

5 of 10

DATASET

Two Fingerprint Datasets, One Metadata Recipe

SOCOFing

6,000 fingerprint images — Model 1

Age

Classes: 18–25, 26–40, 41–64

Height / Weight

Not modeled (NA)

BMI

Under / Normal / Over / Obese classes

PAR

Scale 1–5

FVC2000–DB4–B

814 fingerprint images — Model 2

Age

Uniform, 18–64 years

Height / Weight

150–184 cm / 50–89 kg

BMI

Computed: weight (kg) / height (m)²

PAR

Scale 1–5

Ground truth via Jackson's non-exercise equation: VO₂Max = 56.363 + 10.987·Gender + (1.921·PAR − 0.381·Age − 0.754·BMI)

IECS 2026 · LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING, ANDHRA PRADESH, INDIA

6 of 10

MODEL ARCHITECTURE

CNN + MLP, Fused for a Dual-Task Head

CNN Branch

Modified ResNet-50

1-channel grayscale input

Residual blocks extract ridge features

512-D embedding vector

MLP Branch

Dense-1: 64 neurons, ReLU

Dense-2: 32 neurons, ReLU

Dense-3: 16 neurons, ReLU

Dropout 0.5 to reduce overfitting

Fusion Layer

Z = [ Z_CNN ∥ Z_MLP ]

Concatenated joint embedding

Extra dense layers learn

cross-modal correlations

Dual Output Heads

Regression: linear activation

→ continuous VO₂ Max

Classification: softmax

→ Low / Moderate / High

→

Multitask loss: L_total = α·L_classification + (1−α)·L_regression with α = 0.5, CrossEntropyLoss + HuberLoss, weighted equally

IECS 2026 · LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING, ANDHRA PRADESH, INDIA

7 of 10

RESULTS

Model-1 (SOCOFing) Outperforms Model-2

Metric	Model-1 · SOCOFing	Model-2 · FVC2000–DB4–B
Classification Accuracy	0.8958	0.7937
Precision (Macro)	0.8797	0.7800
Recall (Macro)	0.8918	0.8207
F1-score (Macro)	0.8814	0.7925
MAE (VO₂)	2.08	4.21
Avg. Huber Loss	0.0012	0.0068

Tighter BMI-class binning in Model-1 explains its lower MAE versus raw height/weight in Model-2.

IECS 2026 · LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING, ANDHRA PRADESH, INDIA

8 of 10

RESULTS

Regression Fit & Classification Separability

Predicted vs. actual VO₂ Max — tight clustering along the diagonal

Residual error distribution — centred near zero, few large outliers

Multiclass ROC — AUC of 0.92–1.00 across fitness categories

Results reflect the SOCOFing-fusion (Model-1) run, the strongest performer overall.

IECS 2026 · LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING, ANDHRA PRADESH, INDIA

9 of 10

CONCLUSION

Toward Democratized Preventive Heart Health

A non-invasive hybrid CNN + MLP framework fuses fingerprint biometrics with metadata to estimate VO₂ Max and classify cardiovascular fitness — using Jackson's equation for ground truth, without any lab exercise test.

FUTURE DIRECTIONS

Personalized Coaching

Recommendation engine for exercise guidance and risk alerts

Bigger, Richer Data

More subjects, multiple impressions, minutiae-level ridge features

On-Device Deployment

Mobile & wearable inference via smartphone fingerprint scanners

Privacy-First App

On-device inference with progress tracking and dashboards

IECS 2026 · LAKIREDDY BALI REDDY COLLEGE OF ENGINEERING, ANDHRA PRADESH, INDIA

10 of 10

THANK YOU

Questions & Discussion

Phani Krishna Bulasara

phani.bulasara@gmail.com

Vuyyuru Hema Sri Kanaka Durga

hemasri.v8888@gmail.com

Chinta Prudhvi Sai Raj

chintasai457@gmail.com

Pilli Rupesh

pillirupesh999@gmail.com