Cercar a totes les publicacions

Network meta-analysis and other applications of complex-systems physics to healthcare

Crisafulli,Paride (supervisors: Tobias Galla and Raul Toral)
PhD Thesis (2026)

This thesis presents three distinct applications of computational statistics to
healthcare data: patient flow prediction, comorbidity networks, and network
meta-analysis. Although each application draws on different mathematical
frameworks, they share a common motivation: to extract robust, data-driven
signals from the large volumes of information generated by modern healthcare
to complement the intuition and experience of medical practitioners.
In the first study, we addressed the problem of emergency department over-
crowding, a critical issue for resource allocation and patient safety. Using
seven years of data from the reference hospital of the Balearic Islands (Hospital
Universitari Son Espases, Palma de Mallorca), we developed a non-time-series
machine-learning model to predict the number of incoming patients based
solely on exogenous variables: calendar information, resident population, tourist
arrivals, and weather forecasts. Among random forests, support vector regres-
sors, and feedforward neural networks, the random-forest model proved most
accurate. Counter-intuitively, weather data did not improve predictions since
Mallorca’s mild, seasonal climate was already encoded in calendar variables.
For this reason, the variable was excluded as input. Tourist data, although
statistically significant, yielded an improvement that is small in practice (fewer
than two patients per shift), suggesting that its strong seasonality was largely
captured by calendar information. When compared to a SARIMA time-series
baseline, the random-forest model achieved equivalent predictive accuracy for
horizons of 14 days or longer, offering a viable alternative for long-term resource
planning that does not require a continuous input update. The COVID-19
pandemic permanently altered arrival patterns, highlighting the need to retrain
models when the underlying data-generating process changes.
The second study investigated comorbidity patterns from electronic health
records of more than 600,000 patients in Southwest Finland over 16 years.
Starting from a bipartite patient–disease network, we constructed projected
disease networks and then applied a statistical validation procedure based on the
hypergeometric distribution to retain only the disease pairs whose co-occurrence
was not explainable by chance alone. The resulting Statistically Validated
Networks (SVNs) were substantially sparser than the raw projections and
revealed interpretable community structure when analysed with the Infomap
algorithm. We partitioned patients into cohorts by age and sex, producing 18 dis-
tinct SVNs. Communities of diseases were characterised by over-expression of
specific ICD-10 categories, and a detailed case study of mental and behavioural
disorders illustrated how communities varied with age and sex. A network
dismantling procedure based on betweenness centrality identified which disease
categories most sustained the connectivity of the comorbidity network for each
cohort.
In the third study, we conducted a Bayesian network meta-analysis (NMA) of
aerobic exercise intensity for knee osteoarthritis. In particular, the NMA com-
pared several exercise intensities on two outcomes: pain intensity and walking
vi
performance. Although the analytical framework was designed to identify op-
timal exercise intensity, the results primarily revealed limitations of the existing
literature: few studies per intensity category, predominance of indirect compar-
isons, wide credible intervals, and high risk of bias. The GRADE certainty of
evidence was rated very low for all comparisons, indicating that no reliable con-
clusion about intensity could yet be drawn. Pragmatic clinical recommendations
were offered based on the best available evidence.
The final study systematically compared rank-aggregation methods for network
meta-analysis. We defined the problem in NMA-specific terms, specified de-
sirable properties (handling of input scores, incomplete treatment lists, output
scores, flexible weighting, and the Condorcet criterion), and evaluated the
methods both on synthetic and real data. Two synthetic models were used:
an “all-identical-lists” model to test resilience to missing data, and a model
with correlated, uncorrelated, and anticorrelated outcomes mimicking clinical
datasets. The performance of the methods that used only ordinal ranks and
penalised missing treatments degraded fastest as the fraction of missing treat-
ments increased. Methods that ignored missing data were most resilient but
risked overestimating treatment effects. Network-based methods that accepted
input scores struck a balance between resilience and conservative handling
of missingness. The generalised P-scores method struggled with correlated
outcomes, effectively collapsing blocks of outcomes into a single one. We also
showed that the aggregated rank of placebo could serve as a natural threshold:
if a treatment is ranked below placebo, the strength of its side-effects outweighs
efficacy, given the chosen outcome weights.
Taken together, these studies illustrate how computational statistics can assist
medical decision-making across different scales, from hospital operations to dis-
ease diagnosis to treatment recommendation. The methods and code produced
alongside this thesis are openly available, and the findings are intended to be val-
idated, extended, and applied by other groups to settings beyond those studied
here.

Article complet

Cercar a totes les publicacions

Network meta-analysis and other applications of complex-systems physics to healthcare

Related research projects

CoSASTI

Complex Systems Approach to Social and Technological Interactions

CoSASTI-CSIC

Complex Systems Approach to Social and Technological Interactions

CoSASTI-UIB

Complex Systems Approach to Social and Technological Interactions

APASOS

A Physics approach to sociotechnical systems: from theory to data analysis.

APASOS - CSIC

A Physics approach to sociotechnical systems: from theory to data analysis.

APASOS-UIB

A Physics approach to sociotechnical systems: from theory to data analysis.