A Unified OCR-LSTM Pipeline for Personalized Running Performance Analysis and Forecasting

Abstract

The analysis of running performance is often challenged by data fragmentation across various tracking applications. This study presents a novel mobile application that automates data collection and provides personalized performance forecasting by integrating Optical Character Recognition (OCR) with a Long Short-Term Memory (LSTM) deep learning model. To address data fragmentation, the system allows users to upload running performance summary images, from which an EasyOCR-based module extracts key metrics such as distance and duration. These time-series data are subsequently processed by an LSTM network for two primary tasks: forecasting future running performance over a 30-day horizon and classifying daily activity (run vs. non-run). Experimental results demonstrated the effectiveness of the OCR component, which achieved a low average Word Error Rate (WER) of 10.3%, with 70% of test images recognized perfectly. The LSTM forecasting model yielded a Mean Absolute Error (MAE) of 5.3417 km for distance and 21.1524 minutes for duration. Meanwhile, the binary classification model achieved an accuracy of 68%, with balanced F1-scores of 0.69 (non-running) and 0.67 (running). Furthermore, usability testing confirmed a 100% task completion rate, with automated extraction averaging 18.02 seconds and re-upload times dropping to 8.53 seconds, demonstrating high learnability. These findings validate the feasibility of a unified OCR-LSTM pipeline for running analytics.