A Context-Aware Framework for Biological Health Index Harmonization and Machine Learning-Aided Health Forecasting of Waterbodies
Abstract
National biological monitoring repositories, such as the U.S. EPA’s STORET/WQX, accumulate ecological health scores of water bodies from many agencies using heterogeneous biological indices and scoring conventions. Due to the different standards these agencies adhere to, these scores are not directly comparable. This difference in health score scales and benchmarks undermines large-scale ecological analysis and the application of machine learning models to forecast waterbody health. Existing ML studies predominantly model indices from a single standard framework or utilize physio/chemical parameters for analysis. Studies on composite water quality indices are rare, who also assume homogeneous response variables and regular data sampling. Yet a cursory ook into STORET shows regular data sampling to assess the ecological health of waterbodies is incredibly rare. To develop a harmonization framework for multi-agency biological and habitat indices into a bounded site-level ecological health score, we propose Contextual Temporal Deficit Aggregation (CTDA) that retains assessment context and temporal history. CTDA orients scores, applies rank-based normalization within each index context, performs adaptive temporal smoothing based on the sampling interval, and accumulates persistent deficits before aggregating multiple biological contexts. On a curated national dataset of 17880 index records from 2500 monitoring locations and 38 organizations, CTDA achieved the strongest concordance with independent agency condition classes (Spearman ρ = 0.844), and more than doubled the separation between transient and persistent low-condition events. We then designed an ELM–LSTM hybrid model that combines contextual feature mapping with sequence modelling over CTDA trajectories to forecast health scores. Across rolling forecasting windows, ELM–LSTM outperformed all competing models, achieving the lowest MAE (0.109), RMSE (0.141) and R2 (0.759). By harmonizing diverse biological parameters and coupling them with context-aware sequence forecasting, this study provides a pathway for national-level ecological health trajectory analysis, which boosts targeted management and early warning systems.
Related
- Multiscale harmonization, semantic integration, and immersive exploration of single-cell data in support of novel biological insights
- Imputation-Based Harmonization Mitigates Site Effects Without Data Leakage in Machine Learning Studies
- A heterogeneous biomedical knowledge network framework for rare disease drug candidate prioritization: integrating Orphadata and DisGeNET via gene-bridge harmonization
- Enhancing Strawberry Yield Forecasting With Backcasted IoT Sensor Data and Machine Learning
- Learning and forecasting shared evolutionary pathways to multi-drug resistance across global pathogens
BibTeX
@misc{akter2026contextaware,
title = {{A Context-Aware Framework for Biological Health Index Harmonization and Machine Learning-Aided Health Forecasting of Waterbodies}},
author = {Kazi Fouzia Akter Akter and Sakib Shahriar Shafin and Priyambada Joshi},
howpublished = {EcoEvoRxiv},
year = {2026},
month = oct,
doi = {10.32942/x27d66},
url = {https://ecoevorxiv.org/repository/view/15489/},
note = {CC BY-NC 4.0},
}
RIS
TY - UNPB
TI - A Context-Aware Framework for Biological Health Index Harmonization and Machine Learning-Aided Health Forecasting of Waterbodies
AU - Kazi Fouzia Akter Akter
AU - Sakib Shahriar Shafin
AU - Priyambada Joshi
PY - 2026
DA - 2026/10/29
PB - EcoEvoRxiv
DO - 10.32942/x27d66
UR - https://ecoevorxiv.org/repository/view/15489/
AB - National biological monitoring repositories, such as the U.S. EPA’s STORET/WQX, accumulate ecological health scores of water bodies from many agencies using heterogeneous biological indices and scoring conventions. Due to the different standards these agencies adhere to, these scores are not directly comparable. This difference in health score scales and benchmarks undermines large-scale ecological analysis and the application of machine learning models to forecast waterbody health. Existing ML studies predominantly model indices from a single standard framework or utilize physio/chemical parameters for analysis. Studies on composite water quality indices are rare, who also assume homogeneous response variables and regular data sampling. Yet a cursory ook into STORET shows regular data sampling to assess the ecological health of waterbodies is incredibly rare. To develop a harmonization framework for multi-agency biological and habitat indices into a bounded site-level ecological health score, we propose Contextual Temporal Deficit Aggregation (CTDA) that retains assessment context and temporal history. CTDA orients scores, applies rank-based normalization within each index context, performs adaptive temporal smoothing based on the sampling interval, and accumulates persistent deficits before aggregating multiple biological contexts. On a curated national dataset of 17880 index records from 2500 monitoring locations and 38 organizations, CTDA achieved the strongest concordance with independent agency condition classes (Spearman ρ = 0.844), and more than doubled the separation between transient and persistent low-condition events. We then designed an ELM–LSTM hybrid model that combines contextual feature mapping with sequence modelling over CTDA trajectories to forecast health scores. Across rolling forecasting windows, ELM–LSTM outperformed all competing models, achieving the lowest MAE (0.109), RMSE (0.141) and R2 (0.759). By harmonizing diverse biological parameters and coupling them with context-aware sequence forecasting, this study provides a pathway for national-level ecological health trajectory analysis, which boosts targeted management and early warning systems.
ER -