Modelling and Reasoning in Biomedical Applications with Qualitative Conditional Logic

Jonas Haldimann , Anna Osiak, Christoph Beierle
Knowledge Based Systems
FernUniversität in Hagen, Hagen, Germany

Abstract

  • Three case studies showing how biomedical knowledge can be modelled with conditionals
  • Evaluation of the examples with different established inference methods

Background

Conditional Logic

Syntax

  • Propositional language \(\mathcal{L}\)
  • A conditional: \((A \mid B) \) with \(A, B \in \mathcal{L}\).
  • Set of all conditionals: \((\mathcal{L} \mid \mathcal{L}) = \{ (A \mid B) \mid A, B \in \mathcal{L} \}\).
  • A conditional knowledge base \( \mathcal{R} \subseteq (\mathcal{L} \mid \mathcal{L}) \): A finite set of conditionals.

Semantic

  • Universe \( \Omega \): Set of all models (for the underlying signature)
    E.g., for \(\Sigma = \{a, b\}\) we have \( \Omega_\Sigma = \{ab, a\bar b, \bar a b, \bar a \bar b \} \)
  • A ranking function: \( \kappa: \Omega \rightarrow \mathbb{N}_0\) with \(\kappa^{-1}(0) \neq \emptyset\)
    The most plausible models have rank 0. Less plausible models have higher ranks.
  • Rank of a formula: \(\kappa(A) = \min_{\substack{\omega \in \Omega \\ \omega \models A}} \kappa(\omega)\)
  • A ranking function \(\kappa\) fulfills \((A \mid B) \), denoted \(\kappa \models (A \mid B) \) iff \(\kappa(AB) < \kappa(A\bar B)\).
  • A ranking function \(\kappa\) fulfills a conditional knowledge base \(\mathcal{R}\) iff \(\kappa\) fulfills every conditional in \(\mathcal{R}\).

Inference operators

P-entailment

A knowledge base \(\mathcal{R}\) p-entails \((B \mid A)\) iff every ranking function that fulfills \(\mathcal{R}\) fulfills \(B \mid A\)

System Z

\((B \mid A)\) is a system Z inference if the uniquely determined Pareto-minimal ranking function that fulfills \(\mathcal{R}\) fulfills \((B \mid A)\).

C-inference

  • \((B \mid A)\) is a sceptical c-inference if every c-representation of \(\mathcal{R}\) fulfills \((B \mid A)\).
  • \((B \mid A)\) is a credulous c-inference if at least one c-representation of \(\mathcal{R}\) fulfills \((B \mid A)\).
  • \((B \mid A)\) is a sceptical c-inference if at least one c-representation of \(\mathcal{R}\) fulfills \((B \mid A)\) and no c-representation of \(\mathcal{R}\) does not accept \((B \mid A)\).
  • C-inference can be refined by considering only minimal c-representations.

Modelled Scenarios

Mammals

Scenario

Some information about mammals we would like to model:

  • Mammals cam be divided into three major groups.
  • Most mammals are placentals, i.e., their embryos are nourished by a placenta. Placentals' offspring are born alive (viviparous).
  • Marsupials' are viviparous, but they usually do not develop a (complex) placenta.
  • Monotremes lay eggs. Hence, they do not have a placenta.

Knowledge Base

We use the signature \(\Sigma = \{m, v, c, e, k\}\) where m is true if the animal is a mammal, v if it is viviparous, c if it has a placenta, e if it is a marsupial, and k if it is a monotreme. The resulting knowledge base:

\((v \mid m)\) Mammals are usually viviparous.
\((c \mid m)\) Mammals usually have a placenta.
\((m \mid e)\) Marsupials are mammals.
\((\neg c \mid e)\) Marsupials usually do not have placentas.
\((m \mid k)\) Monotremes are mammals.
\((\neg v \neg c \mid k)\) Monotremes are neither viviparous nor have a placenta.

Evaluation

Query Inf. mode c-inference p-entailment system Z expert opinion
all cw min sum min ind min
\((v \mid e)\) sk. yes yes yes yes no no yes
ws. yes yes yes yes
cr. yes yes yes yes
\((c \mid e)\) sk. no no no no no no no
ws. no no no no
cr. no no no no

Malaria Tropica

Scenario

We want to model the following information about malaria infections.

  • Malaria is caused by an infection with Plasmodium falciparum. But not everyone who is infected gets sick.
  • Patients with the sickle cell allele (a genetic mutation) do not get sick.
  • Infected patients without the sickle cell allele usually get sick.
  • Patients usually do not have the sickle cell allele.
  • Patients with a chemoprophylaxe usually do not get sick, except if infected with a resistant pathogen.

Knowledge Base

We use the signature \(\Sigma = \{m, s, p, r\}\) where m is true if the patient get sick with malaria, s if he has the sickle cell allele, p if he got a chemoprophylaxe, and r if he is infected with a resistant malaria pathogen. We assume that all considered patients are infected with the malaria pathogen.

\((\neg s \mid \top)\) Patients usually do not have the sickle cell allele.
\((m \mid \neg s)\) Infected patients without the sickle cell allele usually get sick.
\((\neg m \mid s)\) Infected patients with the sickle cell allele usually do not get sick.
\((\neg m \mid p)\) Infected patients with a chemoprophylaxis usually do not get sick.
\((m \mid pr)\) Patients with a chemoprophylaxis that are infected with a resistant malaria pathogen usually get sick.

Evaluation

Query Inf. mode c-inference p-entailment system Z expert opinion
all cw min sum min ind min
\((m \mid \top)\) sk. yes yes yes yes yes yes yes
ws. yes yes yes yes
cr. yes yes yes yes
\((m \mid rp)\) sk. yes yes yes yes yes yes yes
ws. yes yes yes yes
cr. yes yes yes yes
\((m \mid rsp)\) sk. no no no no no yes no
ws. no no no no
cr. no no no no

Chronic Myeloid Leukemia

Scenario

Chronic Myeloid Leukemia (CML) is one of the four common forms of leukaemia and is caused by a specific genetic defect.

We want to model the following information.

  • Most cases of CML are caused by a BCR-ABL translocation.
  • Patients with CML caused by the BCR-ABL translocation have good long-term survival rates.
  • Some cases of CML are atypical (aCML). This cases are not caused by a BCR-ABL translocation.
  • aCML can be treated with a hematopoetic stem cell transplantation (HSCT).
  • Recipients of a HSCT can in seldom cases suffer from a severe form of the Graft-versus-Host-Disease (GvHD). Those patients do not have good long-term survival rates.

Knowledge Base

We use the signature \(\Sigma = \{c, a, b, g, m, r\}\) where c is true if the patient has CML, a if he has aCML, b if he has the BCR-ABL translocation, g if he has good chances to survive the CML, m if he gets a HSCT, and r if he suffers from a severe form of GvHD.

\((b \mid c)\) CML is usually caused by a BCR-ABL translocation.
\((g \mid b)\) Patients with a BCR-ABL translocation usually have good survival chances.
\((c \mid a)\) aCML is a form of CML.
\((\neg b \mid a)\) aCML usually coincides with no BCR-ABL translocation.
\((g \mid m)\) Patients getting a HSCT usually have good survival chances.
\((\neg g \mid mr)\) Patients getting a HSCT and suffering from severe GvHD usually have poor survival chances.

Evaluation

Query Inf. mode c-inference p-entailment system Z expert opinion
all cw min sum min ind min
\((m \mid \top)\) sk. yes yes yes yes no yes yes
ws. yes yes yes yes
cr. yes yes yes yes
\((m \mid rsp)\) sk. no no no no no no no
ws. no no no no
cr. no no no no
\((m \mid rp)\) sk. yes yes yes yes no no yes
ws. yes yes yes yes
cr. yes yes yes yes
\((m \mid rsp)\) sk. no no no no no no no
ws. no no no no
cr. no no no no