Causal-graph method maps how COVID-19 symptoms cluster and persist after discharge

Hospital records hold clues about what drives what in a disease, but they are hard to read for cause and effect. A study in PLoS ONE by researchers at Isfahan University of Medical Sciences and the University of Tehran adapts a causal-discovery method to handle that messiness. They apply it to 1,937 patients hospitalised with COVID-19. The authors report that it found clinically coherent clusters of symptoms at admission and persistent links among several symptoms that lingered after discharge. They also report that it beat four rival methods in simulation tests.

Causal discovery tries to work out, from observational data alone, which variables plausibly influence which. Results are often drawn as directed acyclic graphs, or DAGs. These are networks of arrows in which no chain of arrows loops back on itself. Clinical data make this hard. Records mix numbers, such as laboratory values, with categories, such as whether a symptom is present. Measurements unfold over time, values go missing and patients differ widely. Doctors also choose treatments, often according to how sick someone is. That last problem is a form of selection bias. It can make a drug look linked to an outcome for reasons unrelated to the drug.

The team built on an existing framework called Causal Inference over Mixtures (CIM), which represents data as a mixture of several DAGs rather than a single one. They extended it to handle mixed variable types tracked over time. The data came from the Khorshid COVID Cohort and covered admission characteristics, symptom trajectories, in-hospital treatments, laboratory measures and outcomes up to 12 weeks after discharge. The analysis relies on two tools. Conditional-independence tests check whether two variables stay related once other variables are accounted for. Bootstrap stability selection reruns the analysis on many resampled versions of the data and keeps only the links that keep reappearing.

The recovered structures are known as ancestral graphs, a variant of DAG that allows for unmeasured variables. At admission, they showed symptom groupings that the authors describe as clinically coherent. After discharge, several symptoms remained associated with one another, including cough, fatigue or weakness, shortness of breath, weight loss and loss of smell. The authors flag links involving treatments as exploratory. These include the antibiotic ceftriaxone and interferon-β, an immune-signalling protein used as a drug. The authors note that a patient's total exposure to a treatment may simply reflect how ill they were, how long they stayed in hospital or how their doctors made decisions.

To test the method itself, the authors ran simulations in which the true structure is known in advance. They report that CIM consistently outperformed four other causal-discovery algorithms, known as PC, FCI, RFCI and CCI. It achieved the highest recovery accuracy and the highest balanced accuracy, a score that gives equal weight to finding real links and ruling out absent ones. It kept false positives low and was competitive on computing time. This held even when the simulated data included selection bias and hidden confounders, which are unmeasured factors that influence several variables at once.

The authors are explicit about the limits. The data are observational, come from a single cohort and have not been replicated elsewhere. They therefore present the findings as hypothesis-generating rather than as confirmation of causal effects. It remains to be seen whether the symptom links and tentative treatment signals hold in other patient groups. It is also unclear whether the method's advantage in simulation carries over to new real-world datasets.