Melika Baghi

Projects

Research threads

Hospital capacity for medical tourists and local patients

Amirkabir University of Technology · paper under review at Health Care Management Science

arrivals queue Ward A Ward B idle beds shared mean admission queue 25.9 → 14.8 days · 43% shorter
Patients queue for admission; wards lend each other idle beds, cutting the mean queue from 25.9 to 14.8 days.
Problem
A multi-specialty international-patient department serves medical tourists alongside local patients, and how those patients behave (abandoning, switching doctors, choosing a channel) affects the queues.
Approach
A hybrid agent-based and discrete-event simulation in AnyLogic with patient behaviour modelled explicitly, compared against an otherwise identical discrete-event-only model. A compatible-section bed-sharing policy was tested across 30 randomized-seed replications, with a 256-run fractional factorial design screening 16 operational factors.
Result
Medical tourists' hospital-queue wait falls by 82% (13.7 to 2.4 days), and bed sharing cuts the inpatient admission queue by 43% (25.9 to 14.8 days, p < 0.001).
Links
PaperCodeDOI

Healthcare operationsSimulationAnyLogicDesign of experiments

Constrained boosting for reduced-order models

Georgia Institute of Technology

valid unconstrained leaf value outside the valid region projected into both balls so the prediction is always usable enforced at every split, during fit not corrected afterwards
An unconstrained leaf value falls outside the valid region; cXGBoost projects it into the intersection of the balls during fitting.
Problem
Predicting a simulation's reduced basis at unseen parameters means regressing a point on a manifold, valid only inside a bounded region.
Approach
cXGBoost enforces that bound inside every leaf of the boosting ensemble rather than correcting afterwards, so predictions are valid by construction and no general-purpose optimiser is called at any split.
Result
Cuts high-Reynolds reconstruction error 5× against Grassmann interpolation baselines across four PDE problems (cylinder wake, wave propagation, Burgers equation, beam dynamics). Released as an installable open-source package.
Links
PaperCode

Gradient boostingManifold learningSurrogate modeling

Learning from delayed feedback

Georgia Institute of Technology

actions a₁ a₂ a₃ states s₁ s₂ delay d loss at s₁ credited to a₁ and a₂, the actions that reach s₁ price of pooling effective dimension vₜ set by how much actions overlap
Actions lead to states; a loss seen after a delay is credited back only to the actions that could have produced that state.
Problem
When an action's outcome only arrives after a delay, most algorithms wait or discard the signal.
Approach
State-EXP3 charges one delayed outcome to every action that could have produced it, and pays for that pooling with an effective dimension that reads the overlap between actions rather than counting them, so the cost tracks how much the actions actually share.
Result
Regret reduced by up to 79% versus action-level weighting and by 32–68% versus a rate-optimal action-only delayed-bandit baseline, with a lower bound showing the cost of delayed feedback cannot be eliminated by pooling.
Links
PaperCode

Sequential decision makingBanditsOnline learningDelayed feedback

Deciding when partial clinical evidence is enough

Georgia Institute of Technology

observed imaging labs notes genomics enough evidence? yes certify & act no abstain, acquire next coverage holds for every stopping pattern
From the observed modalities, either certify the prediction and act, or abstain and acquire the next modality.
Problem
Clinical models often have to act on incomplete inputs: imaging but no labs, notes but no genomics.
Approach
RouteCert certifies a prediction when the evidence already in hand determines the decision, and otherwise abstains and says which modality to acquire next. Acquisition is separated from calibration, so conformal coverage holds conditional on the terminal modality pattern rather than on average.
Result
Distribution-free coverage guarantees, evaluated on four datasets spanning ECG, speech and wearable sensing (PTB-XL, CMU-MOSEI, MHEALTH, IEMOCAP).
Links
PaperCode

Conformal predictionMultimodal learningMissing dataClinical decision support

Surrogates that keep the data's shape

Georgia Institute of Technology · manuscript under review

simulation field space × space × time one subspace per mode U₁ U₂ U₃ core each mode interpolated on its own manifold cores aligned, then one low-rank surrogate vs flattening the field up to 564× less work 4×–108× faster, same accuracy
Simulation output is multiway; instead of flattening it, each mode keeps its own subspace, interpolated separately, with cores aligned before fitting.
Problem
Simulation output is naturally multiway: space by space by time. The usual move is to flatten it into one long vector, which throws that structure away.
Approach
Every mode gets its own low-dimensional subspace, each interpolated on its own manifold, with the cores aligned before a single low-rank surrogate is fitted.
Result
Across four PDE benchmarks, 35.6× to 564× less interpolation work and 4× to 108× faster than flattened baselines, matching their accuracy to within 0.35 percentage points.
Status
Manuscript under review

Tensor decompositionsSurrogate modelingRiemannian geometryScientific ML

Other projects

Additional projects