A Neu oSymbolic Human-in- he-loop App oach
Towa ds Fusing Medical Expe Knowledge wi h
ANNs
AIAI 2025 - Sil e Tech Wo kshop
Spy os Theodo opoulos∗,∗∗ , Geo gios Mak idis∗∗ , A is odemos
Pne ma ikakiss∗∗∗ , V e os Moulos∗, Dimos henis Ky iazis∗∗ , Panayio is
Tsanakas∗
∗Na ional Technical Uni e si y o A hens
∗∗ Uni e si y o Pi aeus
∗∗∗ Inno a ion Sp in S l
In oduc ion
• The ambi ion o Neu oSymbolic AI (NSAI) is o use symbolic
knowledge de i ed om expe s wi h neu al ne wo ks ha lea n
om as amoun s o da a
• In high-s akes applica ions such as heal hca e, NSAI o e s a
p omising ou e o inc ease anspa ency and us in
au oma ed decision sys ems.
• The managemen o ch onic diseases can se e as a es bed o
NSAI, especially wi h ega ds o gene a ing pe sonalized li es yle
adjus men ad ice.
• Ou main use case is he combina ion o expe guidelines wi h
model knowledge de i ed om Diabe es Melli us Type 2
pa ien s’ li es yle da a, in o de o mi iga e hype glycemia
symp oms and imp o e quali y o li e.
1
The Diabe es Melli us Type 2 Da ase
→Syn he ic da ase p o ided by Inno a ion Sp in , based on ea ly
da a om pa ien s ea ed a he Uni e si y Clinic o
Endoc inology and Me abolic Diseases o he Gene al Uni e si y
Hospi al o La isa.
→Seed da a collec ed using Heal hen ia, ISPs emo e pa ien
moni o ing pla o m om a single pa ien o e he cou se o 8
weeks o syn hesize da a o e he cou se o 14 weeks.
Ac i i y Da a Desc ip ion
S eps Numbe o s eps walked by pa ien on speci ic day
Calo ies Numbe o calo ies consumed
Floo s Amoun o s ai s climbed in loo s
In ensi y minu es Time spen pe o ming in ense exe cise
Sleep The hou s o sleep pe day
Weigh The la es weigh measu emen o he pa ien
2
P edic ion
The ollowing cha ac e is ics o he da ase se he equi emen s o
he NSAI app oach:
•Ta ge : P edic he pa ien ’s SMBG alue using hei ac i i y
cha ac e isics.
• The SMBG alue depends la gely on he ime o he
measu emen e.g. BB: Be o e B eak as .
• P edic ions should be consis en wi h goals se by he doc o s:
Goals:
1. Walk a leas 500 s eps mo e han he baseline
es ablished in weeks -2 and -1.
2. Exe cise a leas 150 mode a ely o in ensi ely (coun ing
x2) minu es pe week.
3. Sleep a leas 6 bu no mo e han 8 hou s e e y nigh .
3
SMBG Values
Figu e 1: Dis ibu ions o SMBG alues pe meal.
4
Neu oSymbolic AI (1/2)
Lea ning o Reasoning s. Reasoning o Lea ning
Lea ning o Reasoning: Ex ensions o symbolic easoning me hods
ha u ilize empi ical machine lea ning o handle uns uc u ed da a
o accele a e easoning.
Reasoning o Lea ning: Symbolic knowledge used in neu al
classi ie s, h ough knowledge ans e o egula iza ion. Key
app oaches include Logic Tenso Ne wo ks (LTN)1and he Symbolic
P obabilis ic Laye (SPL)2.
1Samy Bad eddine e al. “Logic Tenso Ne wo ks”. In: A i icial
In elligence 303 (2022), p. 103649. DOI:
h ps://doi.o g/10.1016/j.a in .2021.103649.
2Ka eem Ahmed e al. “Seman ic P obabilis ic Laye s o
Neu o-Symbolic Lea ning”. In: Ad ances in Neu al In o ma ion P ocessing
Sys ems. Ed. by S. Koyejo e al. Vol. 35. Cu an Associa es, Inc., 2022,
pp. 29944–29959.
5
Neu oSymbolic AI (2/2)
Seman ic P obabilis ic Laye (SPL):
→En o ces logical cons ain s ia O de ed Bina y Decision
Diag ams (OBDDs) ans o med in o di e en iable P obabilis ic
Ci cui s (PCs).
→Readjus s p obabili y con e sions, ensu ing consis ency wi h
ules.
→Low sample complexi y bu limi ed o simple logical
p oposi ions (no i s -o de logic).
Logic Tenso Ne wo ks (LTN):
→G ounding o i s -o de logic p oposi ions o eal- alued
enso s o u h deg ees.
→Tenso s guide back-p opaga ion h ough loss unc ion
egula iza ion.
→High accu acy, low sample complexi y in a ious domains.
→Cons ain s’ sa is iabili y is no ully gua an eed. 6
Towa ds a HIL F amewo k App oach
We p opose a Human-in- he-loop a chi ec u e whe e:
• Medical p o essionals can embed hei expe ise in o a
Neu oSymbolic model by o mula ing human-unde s andable
ules.
• The sys em p o ides explainable AI (XAI) me hods o inspec he
model’s easoning and e alua e co ec ness agains expe
knowledge.
• This enables an i e a i e e inemen p ocess o adjus ules and
imp o e model beha io .
7
HIL F amewo k
Figu e 2: Use case diag am o he p oposed NSAI componen o heal hca e. 8
Conclusion and Fu u e Wo k
• The esul s p esen ed a e a p omising s a ing poin o
combining expe ules wi h da a-d i en lea ning in in elligen
heal hca e applica ions.
• The bene i s o he app oach a e expec ed o become mo e
p onounced on la ge and noisie da ase s, p o ided accu a e
symbolic knowledge is supplied by heal hca e expe s.
•Immedia e ocus: Diabe es use case
• In eg a e anonymized demog aphic and medical examina ion da a
in o ea u e ec o s
• Inco po a e ules de i ed om ques ionnai es and common-sense
guidelines
•Fu u e di ec ion: Explo e me a-cogni i e app oaches whe e ules
a e e ined in e ac i ely wi h use inpu o ensu e alignmen
wi h clinical insigh s.
15
