def _reg_nae_tconst(pattern, qrs):
"""
Temporal constraints for regular beats not coming after ectopic beats.
"""
beats = pattern.evidence[o.QRS]
idx = beats.index(qrs)
assert not _is_ectopic(idx)
hyp = pattern.hypothesis
tnet = pattern.last_tnet
prev = beats[idx - 1]
if idx > 3:
#We create a new temporal network for the new trigeminy cycle.
tnet.remove_constraint(hyp.end, prev.time)
tnet = ConstraintNetwork()
pattern.temporal_constraints.append(tnet)
rrev = beats[idx - 3].time.start - beats[idx - 4].time.start
##RR evolution constraint.
else:
rrev = pattern.evidence[o.Cardiac_Rhythm][0].meas.rr[0]
tnet.add_constraint(prev.time, qrs.time,
Iv(rrev - C.RR_MAX_DIFF, rrev + C.RR_MAX_DIFF))
BASIC_TCONST(pattern, qrs)
tnet.add_constraint(qrs.start, qrs.end, C.NQRS_DUR)
tnet.set_before(qrs.time, hyp.end)
#Constraints with the precedent T Wave
_qrs_after_twave(pattern, qrs)
#Morphology should be similar to the previous QRS, since both are normal
qrs.shape = prev.shape
qrs.paced = prev.paced
def _qrsn_tconst(pattern, qrs):
"""
Temporal constraints for the QRS complexes.
"""
beats = pattern.evidence[o.QRS]
idx = beats.index(qrs)
hyp = pattern.hypothesis
tnet = pattern.last_tnet
obseq = pattern.obs_seq
oidx = pattern.get_step(qrs)
prev = beats[idx-1]
#In cyclic observations, we have to introduce more networks to simplify
#the minimization operation.
tnet.remove_constraint(hyp.end, prev.time)
tnet = ConstraintNetwork()
pattern.temporal_constraints.append(tnet)
meanrr, stdrr = pattern.hypothesis.meas.rr
rr_bounds = Iv(min(C.ASYSTOLE_RR.start, meanrr-stdrr+C.RR_MAX_DIFF),
C.ASYSTOLE_RR.start)
tnet.add_constraint(prev.time, qrs.time, rr_bounds)
tnet.add_constraint(prev.start, qrs.start, rr_bounds)
tnet.add_constraint(prev.end, qrs.end, rr_bounds)
tnet.set_before(prev.end, qrs.start)
#If there is a prior T Wave, it must finish before the start
#of the QRS complex.
if isinstance(obseq[oidx-1], o.TWave):
prevt = obseq[oidx-1]
tnet.set_before(prevt.end, qrs.start)
BASIC_TCONST(pattern, qrs)
tnet.add_constraint(qrs.start, qrs.end, C.QRS_DUR)
tnet.set_before(qrs.time, hyp.end)
#We can introduce constraints on the morphology of the new QRS complex.
if hyp.morph and not qrs.frozen:
qrs.shape = hyp.morph
def _deflection_tconst(pattern, defl):
"""Temporal constraints of the posterior deflections"""
defls = pattern.evidence[o.Deflection]
idx = defls.index(defl)
hyp = pattern.hypothesis
tnet = pattern.last_tnet
prev = defls[idx - 1]
tnet.remove_constraint(hyp.end, prev.time)
#We create a new temporal network for the cyclic observations
tnet = ConstraintNetwork()
tnet.add_constraint(prev.time, defl.time, C.VFLUT_WW_INTERVAL)
pattern.temporal_constraints.append(tnet)
BASIC_TCONST(pattern, defl)
tnet.add_constraint(defl.start, defl.end, Iv(0, C.VFLUT_WW_INTERVAL.end))
tnet.set_before(defl.time, hyp.end)
def _ect_qrs_tconst(pattern, qrs):
"""
Temporal constraints for ectopic beats, which appear after every regular
beat.
"""
beats = pattern.evidence[o.QRS]
idx = beats.index(qrs)
tnet = pattern.last_tnet
hyp = pattern.hypothesis
if idx > 0:
prev = beats[idx - 1]
#After the second couplet, every ectopic beat introduces a new temporal
#network in the pattern to make it easier the minimization.
if idx > 3:
tnet.remove_constraint(hyp.end, prev.time)
#We create a new temporal network for the cyclic observations
tnet = ConstraintNetwork()
pattern.temporal_constraints.append(tnet)
#The duration of each couplet should not have high instantaneous
#variations.
refrr = beats[idx - 2].time.end - beats[idx - 3].time.start
tnet.add_constraint(
prev.time, qrs.time,
Iv(refrr - C.RR_MAX_DIFF, refrr + C.RR_MAX_DIFF))
#We guide the morphology search to be similar to the previous
#ectopic QRS complex.
qrs.shape = beats[idx - 2].shape
#The reference RR varies from an upper limit to the last measurement,
#through the contextual previous rhythm.
refrr = C.BRADY_RR.end
stdrr = 0.1 * refrr
if pattern.evidence[o.Cardiac_Rhythm] and idx == 1:
mrr, srr = pattern.evidence[o.Cardiac_Rhythm][0].meas.rr
if mrr > 0:
refrr, stdrr = mrr, srr
elif idx > 1:
refrr, stdrr = hyp.meas.rr
#Ectopic beats must be advanced wrt the reference RR
tnet.add_constraint(
prev.time, qrs.time,
Iv(C.TACHY_RR.start, max(C.TACHY_RR.start, refrr - stdrr)))
#Beats cannot overlap
tnet.add_constraint(prev.end, qrs.start, Iv(C.TQ_INTERVAL_MIN, np.Inf))
BASIC_TCONST(pattern, qrs)
tnet.add_constraint(qrs.start, qrs.end, C.QRS_DUR)
tnet.set_before(qrs.time, hyp.end)
#Constraints with the precedent T Wave
_qrs_after_twave(pattern, qrs)
def test_add_constraint(self):
# Known example assertion (Detcher STP example in TCN paper)
v0, v1, v2, v3, v4 = [Interval(-np.inf, np.inf) for _ in range(5)]
v0.set(0, 0)
nw = ConstraintNetwork()
nw.add_constraint(v0, v1, Interval(10, 20))
nw.add_constraint(v1, v2, Interval(30, 40))
nw.add_constraint(v3, v2, Interval(10, 20))
nw.add_constraint(v3, v4, Interval(40, 50))
nw.add_constraint(v0, v4, Interval(60, 70))
nw.minimize_network()
assert v0 == Interval(0, 0)
assert v1 == Interval(10, 20)
assert v2 == Interval(40, 50)
assert v3 == Interval(20, 30)
assert v4 == Interval(60, 70)
assert tuple(v0) == (0, 0)
assert tuple(v1) == (10, 20)
assert tuple(v2) == (40, 50)
assert tuple(v3) == (20, 30)
assert tuple(v4) == (60, 70)
# Testing if a stricker constraint is applied
v0, v1, v2, v3, v4 = [Interval(-np.inf, np.inf) for _ in range(5)]
v0.set(0, 0)
nw = ConstraintNetwork()
nw.add_constraint(v0, v1, Interval(10, 20))
nw.add_constraint(v1, v2, Interval(30, 40))
nw.add_constraint(v3, v2, Interval(10, 20))
nw.add_constraint(v3, v4, Interval(40, 50))
nw.add_constraint(v0, v4, Interval(60, 70))
nw.add_constraint(v0, v1, Interval(10, 15))
nw.add_constraint(v1, v2, Interval(30, 35))
nw.add_constraint(v3, v2, Interval(10, 15))
nw.add_constraint(v3, v4, Interval(40, 45))
nw.add_constraint(v0, v4, Interval(60, 65))
nw.minimize_network()
assert v0 == Interval(0, 0)
assert v1 == Interval(10, 10)
assert v2 == Interval(40, 40)
assert v3 == Interval(25, 25)
assert v4 == Interval(65, 65)
def test_add_constraint(self):
# Known example assertion (Detcher STP example in TCN paper)
v0, v1, v2, v3, v4 = [Variable() for _ in range(5)]
v0.value = Interval(0, 0)
nw = ConstraintNetwork()
nw.add_constraint(v0, v1, Interval(10, 20))
nw.add_constraint(v1, v2, Interval(30, 40))
nw.add_constraint(v3, v2, Interval(10, 20))
nw.add_constraint(v3, v4, Interval(40, 50))
nw.add_constraint(v0, v4, Interval(60, 70))
nw.minimize_network()
assert v0.value == Interval(0, 0)
assert v1.value == Interval(10, 20)
assert v2.value == Interval(40, 50)
assert v3.value == Interval(20, 30)
assert v4.value == Interval(60, 70)
# Testing if a stricker constraint is applied
v0, v1, v2, v3, v4 = [Variable() for _ in range(5)]
v0.value = Interval(0, 0)
nw = ConstraintNetwork()
nw.add_constraint(v0, v1, Interval(10, 20))
nw.add_constraint(v1, v2, Interval(30, 40))
nw.add_constraint(v3, v2, Interval(10, 20))
nw.add_constraint(v3, v4, Interval(40, 50))
nw.add_constraint(v0, v4, Interval(60, 70))
nw.add_constraint(v0, v1, Interval(10, 15))
nw.add_constraint(v1, v2, Interval(30, 35))
nw.add_constraint(v3, v2, Interval(10, 15))
nw.add_constraint(v3, v4, Interval(40, 45))
nw.add_constraint(v0, v4, Interval(60, 65))
nw.minimize_network()
assert v0.value == Interval(0, 0)
assert v1.value == Interval(10, 10)
assert v2.value == Interval(40, 40)
assert v3.value == Interval(25, 25)
assert v4.value == Interval(65, 65)
def set_knowledge_base(knowledge):
"""
Sets the knowledge base to be used in the interpretation, and updates
the necessary global variables.
"""
global KNOWLEDGE, _OBSERVABLES, _ABDUCIBLES, _LMAP, _EXCLUSION
KNOWLEDGE = knowledge
#First, we check the consistency of every single abstraction pattern
for p in KNOWLEDGE:
#In the Environment and Abstracted sets no repeated types can happen.
for qset in (p.abstracted, p.environment):
for q in qset:
#The only coincidence must be q
assert len(set(q.mro()) & qset) == 1
#There should be no subclass relations between hyp. and abstractions
for q in p.abstracted:
assert not p.Hypothesis in q.mro()
assert not set(p.Hypothesis.mro()) & p.abstracted
#The abstraction transitions must be properly set.
for q in p.abstractions:
assert q in p.abstracted
for tr in p.abstractions[q]:
assert tr.observable is q
assert tr.abstracted is ABSTRACTED
assert tr in p.transitions
#Organization of all the observables in abstraction levels.
_OBSERVABLES = set.union(*(({p.Hypothesis} | p.abstracted | p.environment)
for p in KNOWLEDGE))
_ABDUCIBLES = tuple(set.union(*(p.abstracted for p in KNOWLEDGE)))
#To perform the level assignment, we use a constraint network.
_CNET = ConstraintNetwork()
#Mapping from observable types to abstraction levels.
_LMAP = {}
for q in _OBSERVABLES:
_LMAP[q] = Variable(value=Interval(0, numpy.inf))
#We set the restrictions, and minimize the network
#All subclasses must have the same level than the superclasses
for q in _OBSERVABLES:
for sup in (set(q.mro()) & _OBSERVABLES) - {q}:
_CNET.set_equal(_LMAP[q], _LMAP[sup])
#Abstractions force a level increasing
for p in KNOWLEDGE:
for qabs in p.abstracted:
_CNET.add_constraint(_LMAP[qabs], _LMAP[p.Hypothesis],
Interval(1, numpy.inf))
#Network minimization
_CNET.minimize_network()
#Now we assign to each observable the minimum of the solutions interval.
for q in _OBSERVABLES:
_LMAP[q] = int(_LMAP[q].start)
#Manual definition of the exclusion relation between observables.
_EXCLUSION = {
Deflection: (Deflection, ),
QRS: (QRS, ),
TWave: (TWave, ),
PWave: (PWave, ),
CardiacCycle: (CardiacCycle, ),
Cardiac_Rhythm: (Cardiac_Rhythm, )
}
#Automatic expansion of the exclusion relation.
for q, qexc in _EXCLUSION.iteritems():
_EXCLUSION[q] = tuple(
(q2 for q2 in _OBSERVABLES if issubclass(q2, qexc)))
for q in sorted(_OBSERVABLES, key=_LMAP.get, reverse=True):
_EXCLUSION[q] = tuple(
set.union(*(set(_EXCLUSION[q2]) for q2 in _EXCLUSION
if issubclass(q, q2))))
def _qrs_tconst(pattern, qrs):
"""
Temporal constraints to observe a new QRS complex.
"""
beats = pattern.evidence[o.QRS]
idx = beats.index(qrs)
hyp = pattern.hypothesis
tnet = pattern.last_tnet
obseq = pattern.obs_seq
oidx = pattern.get_step(qrs)
#The environment complex sets the start of the rhythm observation.
if pattern.get_evidence_type(qrs)[1] is ENVIRONMENT:
tnet.set_equal(hyp.start, qrs.time)
else:
if idx > 0:
prev = beats[idx - 1]
tnet.remove_constraint(hyp.end, prev.time)
#We create a new temporal network for the cyclic observations
tnet = ConstraintNetwork()
tnet.add_constraint(prev.time, qrs.time, rr_bounds)
if rr_bounds is not C.TACHY_RR:
#Also bounding on begin and end, but with relaxed variation
#margin.
rlx_rrb = Iv(rr_bounds.start - C.TMARGIN,
rr_bounds.end + C.TMARGIN)
tnet.add_constraint(prev.start, qrs.start, rlx_rrb)
tnet.add_constraint(prev.end, qrs.end, rlx_rrb)
tnet.set_before(prev.end, qrs.start)
#If there is a prior T Wave, it must finish before the start
#of the QRS complex.
if isinstance(obseq[oidx - 1], o.TWave):
prevt = obseq[oidx - 1]
tnet.set_before(prevt.end, qrs.start)
##RR evolution constraint. We combine the statistical limits
#with a dynamic evolution.
if idx > 1:
prev2 = beats[idx - 2]
rrev = prev.time.start - prev2.time.start
if hyp.meas.rr[0] > 0:
meanrr, stdrr = hyp.meas.rr
const = Iv(
min(0.8 * rrev, rrev - C.RR_MAX_DIFF,
meanrr - 2 * stdrr),
max(1.2 * rrev, rrev + C.RR_MAX_DIFF,
meanrr + 2 * stdrr))
else:
const = Iv(min(0.8 * rrev, rrev - C.RR_MAX_DIFF),
max(1.2 * rrev, rrev + C.RR_MAX_DIFF))
tnet.add_constraint(prev.time, qrs.time, const)
pattern.temporal_constraints.append(tnet)
#TODO improve
if not qrs.frozen and hyp.morph:
nullsh = o.QRSShape()
refbeat = next((b for b in reversed(beats[:idx])
if not b.clustered and all(
b.shape.get(lead, nullsh).tag ==
hyp.morph[lead].tag
for lead in hyp.morph)), None)
if refbeat is not None:
qrs.shape = refbeat.shape
qrs.paced = refbeat.paced
BASIC_TCONST(pattern, qrs)
tnet.add_constraint(qrs.start, qrs.end, C.QRS_DUR)
tnet.set_before(qrs.time, hyp.end)