params.yaml

# Define some default directories, which can then be used
# in the pipeline file `dvc.yaml`.
general:
  data: data/cleaned.csv
  samples: models/samples.hdf5
  risks: models/risks.hdf5
  prevalences: models/prevalences.hdf5
  metrics: metrics.json
  plots_dir: plots

########################################################################################
# This dictionary defines which LNLs are present in the
# graph and how they are connected.
graph:
  tumor:
    primary: [I, II, III, IV, V, VII]
  lnl:
    I: []
    II: []
    III: [IV]
    IV: []
    V: []
    VII: []

########################################################################################
# Below are some parameters necessary for the model setup:
model:
  first_binom_prob: 0.3         # defines shape of time prior of first T-stage
  max_t: 10                     # max number of time steps to evolve system
  t_stages: [early, late]       # T-stages to consider in the data
  class: Unilateral             # model class to use (see `lymph-model` docs)
  # keyword arguments to pass to the model constructor
  kwargs:
    base_symmetric: false
    trans_symmetric: true
    use_mixing: true

########################################################################################
# Choose how to create the synthetic dataset. The currently set values
# reflect what one can see in the data (e.g. on https://lyprox.org)
synthetic:
  t_stages_dist:
    early: 0.6                  # 60% of synethetic patients are early T-stage...
    late: 0.4                   # ...and 40% late T-stage
  midline_ext_prob: 0.3         # create 30% of patients with midline extension

  # Use these made-up specificity & sensitivity values
  # to create the fake observations
  modalities:
    synth_CT: [0.85, 0.85]

########################################################################################
# Under this key we define the specficity and sensitivity for a range
# of diagnostic modalities. They are taken from literature and during
# sampling, everything that's not commented our here will be used for
# inference.
modalities:
  CT: [0.76, 0.81]
  MRI: [0.63, 0.81]
  PET: [0.86, 0.79]
  FNA: [0.98, 0.80]
  diagnostic_consensus: [0.86, 0.81]
  pathology: [1.0, 1.0]
  pCT: [0.86, 0.81]
  max_llh: [1.0, 1.0]

# these of the above modalities will be used to compute things
# like the naive maximum likelihood estimate of the true hidden
# state `max_llh`, or the `rank` "modality".
# The wird nested way of writing them down here is so that DVC can unpack the dict
# directly into the command.
data_cleaning_modalities:
  modalities:
  - CT
  - MRI
  - PET
  - FNA
  - diagnostic_consensus
  - pathology
  - pCT

# this lists the above defined modalities that will be used for
# inference and evaluation.
# The wird nested way of writing them down here is so that DVC can unpack the dict
# directly into the command.
inference_modalities:
  modalities:
  - max_llh

########################################################################################
# This defines the sampler settings and the thermodynamic integration path
sampling:
  walkers_per_dim: 20           # num of parallel walkers per parameter space dimension
  burnin: 1000                  # burn-in steps to discard
  nsteps: 20                    # do this many serious steps per sampling round
  thin_by: 10                   # draw this many samples for one step in `nsteps`
  temp_schedule:                # use this as the inverse temp schedule for TI
  - 0.0
  - 1.0076210988534706e-09
  - 3.224387516331106e-08
  - 2.448519270213934e-07
  - 1.031804005225954e-06
  - 3.148815933917096e-06
  - 7.835261664684588e-06
  - 1.6935087808430282e-05
  - 3.3017728167230525e-05
  - 5.949901826619859e-05
  - 0.00010076210988534707
  - 0.0001622783855914503
  - 0.0002507283732699068
  - 0.0003741226606566017
  - 0.000541922809869769
  - 0.0007651622719418543
  - 0.0010565673013513768
  - 0.0014306778705547923
  - 0.001903968584518355
  - 0.00249496959524998
  - 0.0032243875163311063
  - 0.004115226337448558
  - 0.00519290833892641
  - 0.006485395006257844
  - 0.008023307944637018
  - 0.009840049793490924
  - 0.011971925141011254
  - 0.014458261438686257
  - 0.01734152991583261
  - 0.020667466494127266
  - 0.02448519270213934
  - 0.02884733658986194
  - 0.03381015364324406
  - 0.0394336476987224
  - 0.045781691857753354
  - 0.05292214940134466
  - 0.06092699470458736
  - 0.06987243415118778
  - 0.07983902704799936
  - 0.09091180653955425
  - 0.1031804005225954
  - 0.11673915256060852
  - 0.13168724279835387
  - 0.14812880887639787
  - 0.16617306684564512
  - 0.18593443208187052
  - 0.207532640200251
  - 0.231092867969897
  - 0.2567458542283846
  - 0.28462802079628763
  - 0.31488159339170957
  - 0.34765472254481466
  - 0.38310160451236014
  - 0.4213826021922288
  - 0.46266436603796024
  - 0.5071199549732823
  - 0.5549289573066435
  - 0.6062776116457449
  - 0.6613589278120725
  - 0.7203728077554274
  - 0.7835261664684589
  - 0.8510330529011965
  - 0.923114770875582
  - 1.0
  # only used when sampling until convergence
  kwargs:
    max_steps: 20000
    check_interval: 100
    trust_threshold: 30.0
    rel_acor_threshold: 0.075

########################################################################################
# RISKS:
# Every entry in this array defines a scenario for which the stage
# `predict-risk` will compute the risks, given a set of samples
risks: []


########################################################################################
# PREVALENCES:
# The structure of these scenarios is very similar to the risk.
# But here, one cannot provide a diagnose. This only computes the
# prevalence of a certain pattern of involvement for a defined
# diagnostic modality. It can, however, be compared to the prevalence
# in the data.
prevalences: []