Healthcare Assistance Challenges-Driven Neurosymbolic AI Volume 58- Issue 2

Kaushik Roy

• Student in computer science at the AI Institute, University of South Carolina, USA

Received: July 18, 2024; Published: August 09, 2024

*Corresponding author: Kaushik Roy, student in computer science at the AI Institute, University of South Carolina, Columbia, South Carolina, 29208, USA

DOI: 10.26717/BJSTR.2024.58.009111

Abstract PDF

Although Artificial Intelligence technology has proven effective in providing healthcare assistance by analyzing health data, it still falls short in supporting decision-making. This deficiency largely stems from the predominance of opaque neural networks, particularly in mental health care AI applications, which raise concerns about their unpredictable and unverifiable nature. This skepticism hinders the transition from information support to decision support. This presentation will explore neurosymbolic approaches that combine neural networks with symbolic control and verification mechanisms. These approaches aim to unlock AI’s full potential by enhancing information analysis and decision-making support for healthcare assistance.

Consider the dataset Χ = {(1,1),(2, 4),(3,9)} , in- dexed by , χ_i ^i∈ {1, 2,3} generated from the ground truth function f :χ →χ 2 .Symbolic AI approaches would involve first gathering domain expertise regarding reconstructing the ground truth function, for example, by starting with the set of mappings fˆ :χ →χ , fˆ :χ →χ 2 ,and fˆ :χ →χ 3 , and then letting an expert human map each point in χ_i ∈ Χ. The expert human might map 1 χ = (1,1) to the mapping fˆ :χ →χ ,χ₂ = (2, 4) to the mapping fˆ :χ →χ ² , and χ₃ = (3,9) to the mapping fˆ :χ →χ ² . Subsequently, this mapping can be consolidated to obtain the symbolic approximation as follows:

It is clear from Equation 1 that symbolic AI can lead to wrong approximations due to legitimate human errors (e.g., mapping 1 x = (1,1) to fˆ :χ →χ instead of fˆ :χ →χ ² ). Note that although this approximation is inaccurate, it still generalizes well to all cases of f :χ →χ ² ; Generalization is a powerful benefit of symbolic AI.

Such a function is expected to do well on the training data but generalize poorly too far outside of the training distribution [1]. Furthermore, a significant disadvantage of Neural AI is its black-box nature makes it challenging to interpret and analyze the mappings from inputs to outputs.

Neurosymbolic AI involves using a neural network to first map the data in X to one of the mappings among fˆ :χ →χ , fˆ :χ →χ ² , and fˆ :χ →χ ³ . A neural network will not make the error that the hu- man did in the symbolic AI case, and correctly map (1,1) to f :χ →χ ² in relation to the other data points in X (by minimizing the loss mentioned in the neural AI subsection). Consequently, when such a mapping is consolidated, we obtain an approximation fˆ :χ →χ ² ,which exactly matches the ground truth. This approximation is accurate, does well onthe training data, and generalizes to all χ ∈􀀀 1 . In this way, neurosymbolic AI combines the benefits of both neural and symbolic AI.

Individual Centric Challenges

These challenges are related to the assistance provided by an AI system in individual-specific assessments (e.g., diagnostic or symptom assessments) and recommendations (e.g., treatment methods). The core challenges lie in adequately capturing the individual’s context and adhering to decision-making processes followed by human experts. An Example of Contextualized and Expert Procedure Compliant Decision-Making A person suffering from the most severe form of depressive dis- order must be determined by a PHQ-9 questionnaire- based assessment (specifically, must score in the range of 10-14 on the PHQ-9 assessment scale) (adherence to process followed by expert). Furthermore, the recommended treatment route is pharmacotherapy; however, if the person is found to be exhibiting symptoms of pregnancy, pharmacotherapy is undesirable (contextualized compliance with expert processes).

Overall Infrastructure-Centric Challenges

In the United States, infrastructure issues are one major obstacle to accessing healthcare and implementing AI-driven healthcare support. These include challenges such as healthcare providers struggling to access essential information and protocols within the health- care system. A significant hurdle is creating an AI system that can efficiently navigate the intricate layers of information within the system, helping healthcare professionals streamline care delivery for patients in need. Figure 1 illustrates a concept of AI for tackling infrastructure issues.

Figure 1

Systemic Challenges

For brevity, we refer to the caption in Figure 2 for an explanation of systemic issues such as regulatory compliance, ethical standards, privacy issues, etc.

Figure 2

The talk discusses approaches neuro symbolic AI approaches for handling various challenges in implementing AI-based assistance for improving outcomes. It specifically outlines three levels of challenges at the individual-centric level, the infrastructure- centric level, and the systemic level. I hope this talk will foster innovation at the frontiers of human-AI collaboration for healthcare assistance.

Kaushik Roy is currently pursuing his fourth year as a doctoral student at the AI Institute of the University of South Carolina. His research focuses on developing algorithms that merge statistical data-driven learning techniques with domain-specific information from external sources, with a particular emphasis on ap- plications for social good. He has a track record of publishing numerous papers and presenting talks at prestigious conferences, showcasing his work in areas such as AI for mental health assessment support and other socially impactful applications. The publications venues include IEEE [2–13], NeurIPS [14], NAACL [15, 16], AAAI [17–27], Frontiers [28], KR [29], ECML [30], KDD [31–33], ICAPS [34], KGC [35, 36], ACM [37– 40], ICLR [41], SPIE [42], IJCAI [43], PyData [44], etc. [45–49].

1. Balazs Csanad Csaji (2001) Approximation with artificial neural networks.
2. Prajesh P Anchalia, Kaushik Roy (2014) The k-nearest neighbor algorithm using mapreduce paradigm. In: 5^th International conference on intelligent systems, modelling and simulation, pp. 513-518.
3. Amit Sheth, Kaushik Roy, Manas Gaur (2023) Neu- rosymbolic artificial intelligence (why, what, and how). IEEE Intelligent Systems 38(3): 56-62.
4. Amit Sheth, Manas Gaur, Kaushik Roy, Keyur Faldu (2021) Knowledge-intensive language un- derstanding for explainable ai. IEEE Internet Computing 25(5): 19-24.
5. Amit Sheth, Manas Gaur, Kaushik Roy, Revathy Venkataraman, Vedant Khandelwal, et al. (2022) Process knowledge-infused ai: Toward user-level explainability, interpretability, and safety. IEEE Internet Computing 26(5): 76-84.
6. Azra Nasreen, Kaushik Roy, Kunal Roy, G Shobha (2015) Key frame extraction and foreground modelling using k-means clustering. In 7^th International Conference. on Computational Intelligence, Communication Systems and Networks, pp. 141-145.
7. Manas Gaur, Kaushik Roy, Aditya Sharma, Bi- plav Srivastava, Amit Sheth, et al. (2021) “who can help me?”: Knowledge infused matching of support seekers and support providers during covid- 19 on reddit. In IEEE 9^th International Conference on Healthcare Informatics (ICHI), pp. 265-269.
8. Amit G Bhat, Kaushik Roy, Prajesh P Anchalia, HM Jeevith (2015) Design and implementation of a dynamic intelligent traffic control system. In 17^th UKSim-AMSS International Conference on Modelling and Simulation (UKSim), pp. 369-373.
9. Revathy Venkataramanan, Kaushik Roy, Kanak Raj, Renjith Prasad, Yuxin Zi, et al. (2023) Cook-gen: Ro- bust generative modeling of cooking actions from recipes. In IEEE International Conference on Systems, Man, and Cybernetics (SMC), pp. 981-986.
10. Kaushik Roy, Tarun Garg, Vedant Palit (2023) Knowledge graph guided semantic evaluation of language models for user trust. In IEEE Conference on Artificial Intelligence (CAI), pp. 234-236.
11. Prajesh P Anchalia, Kaushik Roy, Kunal Roy (2015) Two class fisher’s linear discriminant analysis using mapreduce. In 17^th UKSim-AMSS International Conference on Modelling and Simulation (UKSim), pp. 463-469.
12. Amit Sheth, Kaushik Roy (2024) Neurosymbolic value-inspired artificial intelligence (why, what, and how). IEEE Intelligent Systems, 39(1): 5-11.
13. Amit Bhat, Kaushik Roy (2016) Optimized knn for sequential execution. In International Conference on Inventive Computation Technologies (ICICT) 1: 1-6.
14. Kaushik Roy, Vipula Rawte (2022) Tdlr: Top semantic-down syntactic language representation. In NeurIPS’22 Workshop on All Things Attention: Bridging Different Perspectives on Attention.
15. Adam Tsakalidis, Jenny Chim, Iman Munire Bilal, Ayah Zirikly (2022) Overview of the clp sych 2022 shared task: Capturing moments of change in longitudinal user posts. In Proceedings of the Eighth Workshop on Computational Linguistics and Clinical Psychology, pp. 184-198.
16. Shrey Gupta, Anmol Agarwal, Manas Gaur, Kaushik Roy, Vignesh Narayanan, et al. (2022) Learning to automate follow-up question generation using process knowledge for depression triage on reddit posts. In Proceedings of the Eighth Workshop on Computational Linguistics and Clinical Psychology, pp. 137-147.
17. Kaushik Roy, Yuxin Zi, Manas Gaur, Jinendra Malekar, Qi Zhang, et al. (2023) Process knowledge-infused learning for clinician-friendly explanations. In Proceedings of the AAAI Symposium Series 1: 154-160.
18. Kaushik Roy, Usha Lokala, Vedant Khandelwal, Amit P Sheth (2021)” is depression related to cannabis?”: A knowledge-infused model for entity and relation extraction with limited supervision.
19. P Asawa, M Gaur, K Roy, A Sheth (2020) Covid- 19 in Spain and India: Comparing policy implications by analyzing epidemiological and social media data. In 2020 AAAI Fall Symposium on AI for Social Good AI4SG 2020,
20. Kaushik Roy, Vedant Khandelwal, Raxit Goswami, Nathan Dolbir, Jinendra Malekar, et al. (2023) Demo alleviate: Demonstrating artificial intelligence enabled virtual assistance for telehealth: The mental health case. In Proceedings of the AAAI Conference on Artificial Intelligence 37: 16479- 16481.
21. K Roy, Q Zhang, M Gaur, A Sheth (2021) Knowledge infused policy gradients for adaptive pandemic control. In CEUR Workshop Proc.
22. Amit Sheth, Kaushik Roy, Manas Gaur, Usha Lokala (2021) Tutorial on knowledge in-wisdom out-explainable data for ai in cyber social threats and public health.
23. Kaushik Roy, Vedant Khandelwal, Valerie Vera, Harshul Surana, Heather Heckman, et al. (2024) Gear-up: Generative ai and external knowledge-based retrieval Upgrading scholarly article searches for systematic reviews. In Proceedings of the AAAI Conference on Artificial Intelligence 38: 23823-23825.
24. Yuxin Zi, Hariram Veeramani, Kaushik Roy, Amit P Sheth (2023) Rdr: The recap, deliberate, and respond method for enhanced language understanding. In Neuro-Symbolic Learning and Reasoning in the era of Large Language Models.
25. Kaushik Roy, Alessandro Oltramari, Yuxin Zi, Chathurangi Shyalika, Vignesh Narayanan, et al. (2024) Causal event graph-guided language- based spatiotemporal question answering.
26. Yuxin Zi, Kaushik Roy, Vignesh Narayanan, Amit Sheth (2024) Exploring alternative approaches to language modeling for learning from data and knowledge.
27. Kanak Raj, Kaushik Roy, Vamshi Bonagiri, Priyanshul Govil, Krishnaprasad Thirunarayan, et al. (2024) K-perm: Personalized response generation using dynamic knowledge retrieval and persona-adaptive queries.
28. Kaushik Roy, Manas Gaur, Misagh Soltani, Vipula Rawte, Ashwin Kalyan, et al. (2023) Pro- know: Process knowledge for safety constrained and explainable question generation for mental health diagnostic assistance. Frontiers in big Data 5: 1056728.
29. Srijita Das, Sriraam Natarajan, Kaushik Roy, Ron Parr, Kristian Kersting (2020) Fitted q- learning for relational domains.
30. Kaushik Roy, Qi Zhang, Manas Gaur, Amit Sheth (2021) Knowledge infused policy gradients with upper confidence bound for relational bandits. In Machine Learning and Knowledge Discovery in Databases. Research Track: European Conference, ECML PKDD 2021, Bilbao, Spain Proceedings, Part I 21, pp. 35-50.
31. Kaushik Roy, Yuxin Zi, Vignesh Narayanan, Manas Gaur, Amit Sheth (2022) Ksat: Knowledge-infused self-attention transformer- integrating multiple domain-specific contexts.
32. Yuxin Zi, Kaushik Roy, Vignesh Narayanan, Manas Gaur, Amit Sheth (2023) Ierl: Interpretable ensemble representation learning- combining crowdsourced knowledge and dis- tributed semantic representations.
33. Vishal Pallagani, Kaushik Roy, Bharath Muppasani, Francesco Fabiano, Andrea Loreggia, et al. (2024) On the prospects of incorporating large language models (llms) in automated planning and schedul- ing (aps). arXiv preprint arXiv: 2401.02500.
34. Nathan Dolbir, Triyasha Dastidar, Kaushik Roy (2021) Nlp is not enough-contextualization of user input in chatbots.
35. Kaushik Roy, Manas Gaur, Qi Zhang, Amit Sheth (2022) Knowledge-infused reinforcement learning.
36. Kaushik Roy, Usha Lokala, Manas Gaur, Amit P Sheth (2022) Tutorial: Neuro-symbolic ai for mental healthcare. In Proceedings of the Second International Conference on AI-ML Systems, p. 1-3.
37. Valeria Fionda, Olaf Hartig, Reyhaneh Abdolazimi, Sihem Amer-Yahia, Hongzhi Chen, et al. (2023) Tutorials at the web conference 2023. In Companion Proceedings of the ACM Web Conference, pp. 648-658.
38. Aidin Shiri, Kaushik Roy, Amit Sheth, Manas Gaur (2023) L3 ensembles: Lifelong learning approach for ensemble of foundational language models.
39. Usha Lokala, Kaushik Roy (2023) Tutorial: Causal ai for web and health care. In Companion Proceedings of the ACM Web Conference.
40. Usha Lokala, Francois Lamy, Triyasha Ghosh Dastidar, Kaushik Roy, Raminta Daniulaityte, et al. (2021) Edark- trends: Harnessing social media trends in sub- stance use disorders for opioid listings on cryp- tomarket. In ICLR AI for Public Health Workshop.
41. Sandeep K Chaudhuri, Joshua W Kleppinger, Ritwik Nag, Kaushik Roy, Rojina Panta, et al. (2021) A cdzntese gamma spectrometer trained by deep convolu- tional neural network for radioisotope identification. In Hard X-Ray, Gamma-Ray, and Neutron Detector Physics XXIII 11838, pp. 1183806.
42. Usha Lokala, Manas Gaur, Kaushik Roy, Amit Sheth (2021) Knowledge infused natural language understanding for public health, epidemiology, and substance use.
43. Kaushik Roy, Manas Gaur (2020) Knowledge- infused statistical learning for social good.
44. Chathurangi Shyalika, Kaushik Roy, Renjith Prasad, Fadi El Kalach, Yuxin Zi, et al. (2024) Ri2ap: Robust and interpretable 2D anomaly prediction in assembly pipelines.
45. Kaushik Roy (2024) Tutorial: Knowledge-infused artificial intelligence for mental healthcare.
46. Abhisek Tiwari, Muhammed Sinan, Kaushik Roy, Amit Sheth, Sriparna Saha, et al. (2023) Hi model, generating’nice’instead of’good’is not as bad as generating’rice’! towards context and semantic infused dialogue generation loss function and evaluation metric. arXiv preprint arXiv: 2309.05804.
47. Kaushik Roy (2023) Tutorial-shodhguru labs: Knowledge-infused artificial intelligence for mental healthcare.
48. Kaushik Roy (2023) Tutorial-shodhguru labs: Optimization and hyperparameter tuning for neural networks.
49. Kaushik Roy (2020) Relational sequential decision making.