Interpretable neural networks based on continuous-valued logic and multicriteria decision operators

Csiszár, Orsolya; Csiszár, Gábor; Dombi, József

doi:10.1016/j.knosys.2020.105972

Cited by 34 publications

(20 citation statements)

References 15 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…Our strategy consists of implementing networks based on logical gates, modeled by Perceptron with fixed weights and biases. This hybrid neural model was introduced in [ 9 , 10 ]. Here, a single Perceptron in the NN network is activated by so-called Squashing activation functions, differentiable, a parametric family of functions that satisfy natural invariance requirements and contain rectified linear units as a particular case [ 16 , 17 ].…”

Section: Methodsmentioning

confidence: 99%

Medical recommender systems based on continuous-valued logic and multi-criteria decision operators, using interpretable neural networks

Ochoa¹,

Csiszár

Schimper³

2021

BMC Med Inform Decis Mak

Self Cite

View full text Add to dashboard Cite

Background Out of the pressure of Digital Transformation, the major industrial domains are using advanced and efficient digital technologies to implement processes that are applied on a daily basis. Unfortunately, this still does not happen in the same way in the medical domain. For this reason, doctors usually do not have the time or knowledge to evaluate all alternative treatment options for each patient accurately and individually. However, physicians can reduce their workload by using recommender systems, still having every decision under control. In this way, they also get an insight into how other physicians make treatment decisions in each situation. In this work, we report the development of a novel recommender system that uses predicted outcomes based on continuous-valued logic and multi-criteria decision operators. The advantage of this methodology is that it is transparent, since the model outcomes emulate logical decision processes based on the hierarchy of relevant physiological parameters, and second, it is safer against adversarial attacks than conventional deep learning methods since it drastically reduces the number of trainable parameters. Methods We test our methodology in a patient population with diabetes and heart insufficiency that becomes a therapy (beta-blockers, ACE or Aspirin). The original database (Pakistan database) is publicly available and accessible via the internet. However, to explore methods to protect the patient's identity and guarantee data privacy we implemented a methodology on a variable-by-variable basis by fitting a sequence of regression models and drawing synthetic values from the corresponding predictive distributions using linear regressions and norm rank. Furthermore, we implemented a deep-learning model based on logical gates modeled by perceptrons with fixed weights and biases. While a first trainable layer automatically recognizes a meaningful parameter hierarchy, the implemented Logic-Operator Neuronal Network (LONN) simulates cognitive processes like a rational, logical thinking process, considering that this logic is joined by fuzziness, i.e., logical operations are not exact but essentially fuzzy due to the implemented continuous-valued operators. The predicted outcomes of the model (kind of therapy-ACE, Aspirin or beta-blocker- and expected therapy time of the patient) are then implemented in a recommender system that compares two different models: model 1 trained on a population excluding negative outcomes (patient group 1, with no patient dead and long therapy times) and a model 2 trained on the whole patient population (patient group 2). In this way, we provide a recommendation of the best possible therapy based on the outcome of the model and the confidence of this recommendation when the outcome of model 1 is compared with the outcome of model 2. Results With the applied method for data synthetization, we obtained an error of about 1% for all the relevant parameters. Furthermore, we demonstrate that the LONN models reach an accuracy of about 75%. After comparing the LONN models against conventional deep-learning models we observe that our implemented models are less accurate (accuracy loss of about 8%). However, the loss of accuracy is compensated by the fact that LONN models are transparent and safe because the freezing of training parameters makes them less prone to adversarial attacks. Finally, we predict the best therapy as well as the expected therapy time. We were able to predict individualized therapies, which were classified as optimal (binary value) when the prediction fully matched predictions made with models 1 and 2. The results provided by the recommender system are displayed using a graphical interface. The current is a proof of concept to improve the quality of the disease management, while the methods are continuously visualized to preserve transparency for the customers. Conclusions This work contributes to simplify administrative functions and boost the quality of management of patients improving the quality of healthcare with models that are both transparent and safe. Our methodology can be extended to different clinical scenarios where recommender systems can be applied. The acceptance and further development of the app is one of the next important steps and still requires further development depending on specific requirements of the health management, the physicians or health professionals, and the patent population.

show abstract

Section: Methodsmentioning

confidence: 99%

Medical recommender systems based on continuous-valued logic and multi-criteria decision operators, using interpretable neural networks

Ochoa¹,

Csiszár

Schimper³

2021

BMC Med Inform Decis Mak

Self Cite

View full text Add to dashboard Cite

show abstract

“…Preliminary analysis shows that for some applications, it is more advantageous to use different activation functions for different neurons -i.e., select a family of activation functions instead, and select the parameters of activation functions of different neurons during training. Specifically, this was shown for a special family of squashing activation functions that contain rectified linear neurons as a particular case; see, e.g., [2]- [4]. Functions from this family have the form…”

Section: Shall We Go Beyond Rectified Linear Activation Functions?mentioning

confidence: 98%

Why Squashing Functions in Multi-Layer Neural Networks

Urenda

Csiszár

et al. 2020

2020 IEEE International Conference on Systems, Man, and Cybernetics (SMC)

View full text Add to dashboard Cite

“…Our strategy consists of implementing networks based on logical gates, modeled by Perceptron with fixed weights and biases. This hybrid neural model was introduced in (Csiszár et al, 2020a). Here, a single Perceptron in the NN network is activated by so-called Squashing activation functions, differentiable, parametric family of functions that satisfy natural invariance requirements and contain rectified linear units as a particular case (Urenda et al, 2020a;Zeltner et al, 2020).…”

Section: Continuous-valued Logic Multi-criteria Decision Operators Anmentioning

confidence: 99%

“…Thus, the Perceptron in the neural networks' hidden layers can model a threshold-based nilpotent operator (Csiszár et al, 2020a(Csiszár et al, , 2020b): a conjunction, a disjunction, or even an aggregative operator. This means that the weights of the first layer is to be learned, while the hidden layers of the pre-designed neural block, worked as logical operators with frozen weights and biases.…”

Section: Continuous-valued Logic Multi-criteria Decision Operators Anmentioning

confidence: 99%

“…Here, we are introducing a 'different recommender system' which is a combination of methods to generate synthetic populations while keeping personal data protected when needed for Artificial Intelligence (AI) applications and the use of novel methods based on continuous-valued logic and multi-criteria decision operators aimed for a robust, safer, and more understandable use of deep learning. By combining neural networks with continuous logic and multi-criteria decision-making tools, thus reducing in this way the black-box nature of neural models (Csiszár et al, 2020a(Csiszár et al, , 2020bRiegel et al, 2020;Shi et al, 2019). By doing this, we are exploring the (often overlooked) possibility of combining neural networks with continuous logical systems.…”

Section: Introductionmentioning

confidence: 99%

See 1 more Smart Citation

Medical Recommender Systems based on continuous-valued logic and multi-criteria decision operators, using interpretable neural networks

Ochoa¹,

Csiszár

Schimper³

2021

Preprint

Self Cite

View full text Add to dashboard Cite

Background Out of the pressure of Digital Transformation, the major industrial domains are using advanced and efficient digital technologies to implement processes that are applied on a daily basis. Unfortunately, this still does not happen in the same way in the medical domain. For this reason, doctors usually do not have the time or knowledge to evaluate all alternative treatment options for an individual patient accurately and individually. However, physicians can reduce their workload by using Recommender Systems, still having every decision under control. In this way, they also get an insight into how other physicians make treatment decisions in each situation. In this work, we report the development of a novel recommender system that uses predicted outcomes based on continuous-valued logic and multi-criteria decision operators. The advantage of this methodology is that it is transparent, since the model outcomes emulate logical decision processes based on the hierarchy of relevant physiological parameters, and second, it is safer against adversarial attacks than conventional deep learning methods since it drastically reduces the number of trainable parameters.Methods We test our methodology in a patient population with diabetes and heart insufficiency that becomes a therapy (beta-blockers, ACE or Aspirin). The original database (Pakistan database) is publicly available and accessible via the internet. However, to explore methods to protect the patient`s identity and guarantee data privacy we implemented a methodology on a variable-by-variable basis by ﬁtting a sequence of regression models and drawing synthetic values from the corresponding predictive distributions using linear regressions and norm rank.Furthermore, we implemented a deep-learning model based on logical gates modeled by perceptrons with fixed weights and biases. While a first trainable layer automatically recognizes a meaningful parameter hierarchy, the implemented Logic-Operator Neuronal Network (LONN) simulates cognitive processes like a rational, logical thinking process, considering that this logic is joined by fuzziness, i.e., logical operations are not exact but essentially fuzzy due to the implemented continuous-valued operators.The predicted outcomes of the model (kind of therapy -ACE, Aspirin or beta-blocker- and expected therapy time of the patient) are then implemented in a recommender system that compares two different models: model 1 trained on a population excluding negative outcomes (patient group 1, with no patient dead and long therapy times) and a model 2 trained on the whole patient population (patient group 2). In this way, we provide a recommendation of the best possible therapy based on the outcome of model 1and the confidence of this recommendation when the outcome of model 1 is compared with the outcome of model 2. Results With the applied method for data synthetization, we obtained an error of about 1% for all the relevant parameters. Furthermore, we demonstrate that the LONN models reach an accuracy of about 75%. After comparing the LONN models against conventional deep-learning models we observe that our implemented models are less accurate (accuracy loss of about 8%). However, the loss of accuracy is compensated by the fact that LONN models are transparent and safe because the freezing of training parameters makes them less prone to adversarial attacks. Finally, we predict the best therapy as well as the expected therapy time. We were able to predict individualized therapies, which were classified as optimal (binary value) when the prediction fully matched predictions made with models 1 and 2. The results provided by the recommender system are displayed using a graphical interface. The current is a proof of concept to improve the quality of the disease management, while the methods are continuously visualized to preserve transparency for the customers. Conclusions This work contributes to boost administrative functions and quality of management of patients and improve the quality of healthcare with models that are both transparent and safe. Our methodology can be extended to different clinical scenarios where recommender systems can be applied. The acceptance and further development of the app is one of the next important steps and still requires further development depending on specific requirements of the health management, the physicians or health professionals, and the patent population.

show abstract

Interpretable neural networks based on continuous-valued logic and multicriteria decision operators

Cited by 34 publications

References 15 publications

Medical recommender systems based on continuous-valued logic and multi-criteria decision operators, using interpretable neural networks

Medical recommender systems based on continuous-valued logic and multi-criteria decision operators, using interpretable neural networks

Why Squashing Functions in Multi-Layer Neural Networks

Medical Recommender Systems based on continuous-valued logic and multi-criteria decision operators, using interpretable neural networks

Contact Info

Product

Resources

About