1. River assessments are predominantly based upon biological metrics and indices selected or designed to integrate the impact of multiple causes of deterioration (stressors) operating at various spatial scales. Yet, the integrative nature of many bioassessment systems does not allow for tracing back individual stressors and their influence on the overall assessment result. Thus, river managers often fail to link bioassessment with programmes of management measures, to improve ecological quality.
2. Here, we present a novel diagnostic approach that allows to estimate the probability of individual stressors being causal for biological degradation at the scale of individual riverine ecosystems. Similar to medical diagnosis, we use various symptoms (macroinvertebrate metrics) and probabilistically link them to various potential causes of ecological status degradation (stressors). Symptoms and causes are informed by a training dataset of 157 samples (stressors, taxa lists) from central European lowland rivers and are linked through a Bayesian Network (BN). Three separate BNs addressing three different spatial scales (catchment, reach and site) are presented. 
3. Water quality-related causes are most influential at the catchment scale, while hydromorphological causes prevail at finer scales. Causes indicating riparian degradation are most influential at the reach scale. Many symptoms show strong linkages to causes and reveal ecologically meaningful relationships, thus pointing at the potential diagnostic utility of the symptoms selected. BNs are validated using an independent dataset of 47 samples. Overall, model accuracies range 53–58% for the three BNs, while for individual nodes (causes and symptoms) up to 100% concordance of predicted and actual node states in the validation data is achieved. The BNs are implemented as interactive online diagnostic tools to allow end users an easy application. 
4. Synthesis and applications. Our results confirm that Bayesian inference can greatly assist the diagnosis of potential causes of river deterioration based upon a selection of diagnostic biological metrics. If integrated into a Bayesian Network, symptoms and potential causes can be linked and inform management decisions on appropriate measures, to improve ecological quality. Diagnostic Bayesian Networks thus support end users bridge the gap between biological monitoring and appropriate programmes of management measures. 28-Jul-2020

The dataset consists of river benthic macroinvertebrate metrics and accompanying environmental covariates sampled between 2000 and 2005 from sites in the central European lowlands. Macroinvertebrate metrics encompass assessment indices, diversity measures and traits (e.g. feeding types). The metrics were calculated from taxalists (mainly species-level identification) using the traits and ecological characteristics provided by www.freshwaterecology.info. Macroinvertebrate taxa lists were obtained by multi-habitat sampling at 50–100 m long river stretches using a hand net (frame 50 x 50 cm, mesh: 500 µm). Samples were preserved in ethanol (70% end concentration) and stored for determination in the lab. 

Environmental covariates encompass physical-chemical, hydrological, morphological (incl. microhabitats) and land cover variables at various levels of intensity, i.e at different levels of human-induced stress imposed on the riverine macroinvertebrate community. Site and reach-scaled variables (physical-chemical, morphological, hydrological) were recorded in parallel to biological sampling, while land cover and other catchment-scale data were obtained through maps in a GIS.

Field sampling methodologies and lab procedures including the chemical analysis of water samples, the determination of macroinvertebrates and the calculation of metrics are described in detail in two publications: 
- Feld, C.K. (2004) Identification and measure of hydromorphological degradation in Central European lowland streams. Hydrobiologia, 516, 69–90.
- Feld, C.K. & Hering, D. (2007) Community structure or function: effects of environmental stress on benthic macroinvertebrates at different spatial scales. Freshwater Biology, 52, 1380–1399.

The dataset is provided as an Excel file. The first table contains the explanation of column codes, the variables units and the references to the variables, if available.

The second table contains the matrix of 144 samples (rows) by environmental variables and biological metrics (columns) used to develop and train the Bayesian Networks.

The third table contains a similar matrix of 47 independent samples and environmental/biological variables used to validate the Bayesian Networks. This validation set contains a number of missing values (coded NA). 

Please note that the data contain original measurements and records, respectively, which may require discretisation, if to be used for the training and validation of a discretised Bayesian Network.

For further details, please contact christian.feld@uni-due.de.