HealthProcessAI: AI in Healthcare Processes

• HealthProcessAI is a process-centric framework that integrates AI into sequential care delivery, covering intake, triage, diagnosis support, and chronic management.
• It leverages methodologies like process mining, multimodal modeling, and sequential decision-making to optimize healthcare workflows and enhance diagnostic accuracy.
• The framework addresses key challenges such as drift, governance, and scalability, utilizing modular components, ontology-driven insights, and real-time audit protocols.

Searching arXiv for the explicit "HealthProcessAI" framework and closely related healthcare-process AI papers to ground the article in current arXiv literature. {"query":"HealthProcessAI healthcare process AI process mining triage arXiv", "max_results": 10} arXiv search results were consulted to verify the explicit "HealthProcessAI" framework and neighboring literature on process mining, triage, and AI-enabled healthcare workflows, including (Illueca-Fernandez et al., 29 Aug 2025, Marchiori et al., 2020), and related healthcare-process AI papers. HealthProcessAI designates a class of systems in which artificial intelligence is embedded in healthcare processes rather than confined to isolated prediction tasks. In the literature, the term appears explicitly as a GenAI framework that wraps PM4PY and bupaR with multiple LLMs for healthcare process mining, and it is also used more broadly to describe AI woven into the operational fabric of healthcare, beginning at intake and triage and extending through documentation, diagnostics, follow-up, chronic disease management, public-health monitoring, and process governance (Illueca-Fernandez et al., 29 Aug 2025, Marchiori et al., 2020). Early instantiations include questionnaire-driven symptom assessors on mobile devices, while later systems couple ontologies, knowledge graphs, multimodal models, federated architectures, and process-aware interfaces to support both clinical and organizational workflows (Khan et al., 2018, Chen et al., 23 Feb 2025).

1. Conceptual foundations

HealthProcessAI is process-centric. Its defining concern is not merely whether an algorithm predicts a label, but how AI participates in the ordered sequence of activities by which care is delivered: symptom elicitation, triage, referral, diagnosis support, treatment recommendation, adherence support, follow-up, and workflow analysis. In the triage literature, this is stated directly: AI operates “at the very beginning” of the chain of care, becoming the “front door” of telemedicine and shaping which patients end up where, and when (Marchiori et al., 2020).

A second foundational theme is the distinction between domain knowledge and process knowledge. “Process knowledge” is defined as “an ordered set of information that maps to evidence-based guidelines or categories of conceptual understanding to experts in a domain” (Sheth et al., 2022). In HealthProcessAI, this ordered structure may be a questionnaire, a rule tree, a guideline pathway, a process map, or a metamodel of workflows. The implication is that safe healthcare AI must do more than recognize patterns; it must also follow clinically meaningful sequences of actions.

A third theme is scoping. The clinical-AI viewpoint literature argues that AI is strongest on narrow, well-defined questions and should be matched to the workflow step, the available data, and the validation regime. That argument is crystallized in the 6R framework: relevant clinical question, right data, appropriate ratio between number of patients and variables, relationship between data and ground truth as direct and causal as possible, regulatory ready validation, and right AI method (Hartskamp et al., 2019). This suggests that HealthProcessAI is best understood as a modular assemblage of process-aligned components rather than a single monolithic “doctor model.”

2. Representations of healthcare processes

The simplest HealthProcessAI representation in the supplied literature is the symbolic decision tree. In the Android “smart doctor” system, each question is an internal node, each answer selects a branch, and each leaf corresponds to a treatment decision. The tree is manually constructed from medical knowledge, and the dialogue path is the diagnostic path itself (Khan et al., 2018). In formal terms, this is a deterministic questionnaire process whose control structure is explicit and inspectable.

More elaborate systems replace hand-built trees with ontological and graph-based representations. The Artificial Intelligence Triage Engine (AITE) uses a multi-stage NLP pipeline to extract symptoms, diseases, negations, and relations from more than 900,000 German-language teleconsultation records, then clusters roughly 1 million raw expressions into 17,000 medical concepts and 100,000 relations. These concepts populate an ontology and knowledge graph used for both question generation and recommendation (Marchiori et al., 2020). Here, process structure is distributed across concept hierarchies, red-flag annotations, patient-similarity subgraphs, and triage endpoints.

Model-driven work pushes process representation even further upward, into explicit metamodels and domain-specific languages. The MDE4AI framework introduces MILA, a healthcare-specific DSL in which workflows, tasks, cohorts, data elements, and federation policies are specified at a platform-independent level, validated against ontologies, transformed into platform-specific models, and then compiled into executable code (Raheem et al., 10 Oct 2025). In this form, HealthProcessAI is not only a runtime system but a software-engineering discipline for healthcare process specification.

Process mining contributes another representational layer. The explicit “HealthProcessAI” framework models healthcare data as event logs composed of cases, activities, timestamps, and optional clinical attributes, then derives directly-follows graphs, Petri-net-like structures, and narrative reports from those logs (Illueca-Fernandez et al., 29 Aug 2025). In this setting, the process is not prescribed in advance; it is discovered from observed traces.

Dialogue-based pre-consultation systems introduce a centralized state representation. The multi-agent pre-consultation framework decomposes intake into Triage, History of Present Illness collection, Past History collection, and Chief Complaint generation, with 13 domain-specific subtasks. A “Recipient” agent maintains evolving HPI, PH, and CC states across turns, while a “Monitor” scores subtask completion and a “Controller” schedules the next subtask (Yu et al., 3 Nov 2025). This is a process representation tailored to long, context-sensitive conversations.

3. Methodological repertoire

HealthProcessAI spans symbolic, statistical, neural, graph-based, and sequential-decision methods. At the symbolic end, manually engineered decision trees implement transparent top-down search over symptom attributes; this is the operative logic in the smartphone “smart doctor” prototype (Khan et al., 2018). At the opposite end, telemedicine triage uses dictionary-based NLP, rules, CNN/Bi-GRU/Bi-LSTM relation extraction, query expansion metrics, neural question generation, graph embeddings, and multi-class classifiers linked to risk and point-of-care decisions (Marchiori et al., 2020).

A central methodological distinction is between static inference and sequential decision-making. The Markov decision process framework for simulated clinical decision-making models patient evolution with transition probabilities $TR(s_{t+1}, s_t, a_t)=P(s_{t+1}\mid s_t,a_t)$ and, under partial observability, updates belief states using $b_t^o(s_t)=\frac{P(o_t\mid s_t)b_t(s_t)}{Z}$. The system then plans and re-plans over visit sequences while optimizing Cost per Unit Change (Bennett et al., 2013). This is particularly relevant to HealthProcessAI because many healthcare processes are inherently longitudinal and policy-dependent.

Multimodal modeling is another recurrent methodology. The HAIM framework builds tabular, time-series, text, and image embeddings and concatenates them into a fusion embedding, training 14,324 models across 12 predictive tasks; multimodal configurations outperform comparable single-source approaches by 6–33% (Soenksen et al., 2022). DeepSeek deployments in tertiary hospitals describe a LLM core extended with vision, retrieval, and workflow-assistance modules, with hierarchical knowledge distillation reducing inference cost by about 30% (Chen et al., 23 Feb 2025). A plausible implication is that HealthProcessAI increasingly relies on modality-specific encoders feeding shared workflow interfaces.

For chronic disease and personalized medicine, the literature combines ontology-driven preprocessing with task-specific predictors: cluster analysis for diabetes subtypes, graph neural networks for treatment response, support vector machines for retinopathy onset, transformer-style sequence modeling for cardiovascular events, federated learning for multi-center collaboration, and Shapley-based explanation layers (Maranghi et al., 2022, Ivanovic et al., 2022). In public-health surveillance, the dominant methodological move is away from fixed thresholds and toward ranking-based anomaly scoring: every data point receives a continuous score and reviewers triage a ranked list rather than a binary alert queue (Joshi et al., 4 Jun 2025).

Across these methods, explainability is repeatedly tied to process structure rather than to post hoc visualization alone. Process knowledge-infused AI represents a workflow as an ordered sequence $P=(s_1,\dots,s_n)$ or as a directed graph $G=(V,E)$ and constrains generation, questioning, or recommendation so that outputs remain inside expert-defined pathways (Sheth et al., 2022). This suggests that in HealthProcessAI, interpretability is often an emergent property of explicit process control.

4. Principal application domains

The most immediate HealthProcessAI domain is pre-consultation. The Android smart-doctor system acts as a personal medical assistant on a smartphone, performs initial triage for common conditions, produces diagnosis-like statements, recommends medicines such as Desprine, Brufen, Flagyl, Augmentin, and ORS, and sets alarms for medication timing (Khan et al., 2018). AITE extends the same front-door logic to telemedicine, where the mobile app asks AI-powered personalized questions and routes patients toward immediate teleconsultation, scheduled teleconsultation, GP visit, or emergency care (Marchiori et al., 2020). The multi-agent pre-consultation architecture further formalizes intake into triage, HPI, PH, and chief-complaint generation, turning passive symptom chat into proactive inquiry (Yu et al., 3 Nov 2025).

Hospital workflow support is a second major domain. DeepSeek deployments across China’s tertiary hospitals include pathology annotation and Ki-67 quantification at Ruijin Hospital, imaging-plus-biomarker diagnosis at Huashan Hospital, AI drafting of medical records at Shanghai Fourth People’s Hospital, physician workstation integration and pre-consultation at the Jinshan Branch of Shanghai Sixth People’s Hospital, evidence retrieval and personalized treatment recommendations in urology at South China Hospital, hematology and lab AI at Liuzhou People’s Hospital, and chronic disease management plus telemedicine at Chengdu First People’s Hospital (Chen et al., 23 Feb 2025). The Egyptian AI-driven EHR system expresses a related architecture at the infrastructure level: microservices, polyglot persistence, Llama3-OpenBioLLM-70B for summarization and chatbot functions, and a Vision Transformer for pneumonia classification, all embedded into patient examination and history tracking workflows (Alorbany et al., 8 Feb 2025).

Chronic disease management is a third domain. The AMD diabetes project uses ontology-based conceptualization of 600,000 patients with type 2 diabetes across about 320 centers to support subgroup discovery, disease progression modeling, retinopathy prediction, HbA1c trajectory clustering, and short-term cardiovascular risk forecasting (Maranghi et al., 2022). The personalized-medicine survey describes a broader pipeline in which Holistic Health Records combine EHR, QoL, wearable, sensor, environmental, and lifestyle data, while federated learning and homomorphic encryption support privacy-preserving prediction and intervention simulation (Ivanovic et al., 2022). User-centric work generalizes the same process view to diabetes, obesity, stress, dyslexia, and cancer by treating chronic conditions as ongoing, highly individualized processes rather than static diagnoses (Ayesh, 2024).

HealthProcessAI also extends beyond direct clinical encounters. The AI-enabled public-health monitoring system ranks anomalies across up to 5,000,000 data points daily so that reviewers can triage outbreaks and data-quality issues faster than with threshold alerts (Joshi et al., 4 Jun 2025). The AI4HCTI platform applies process-aware AI to cyber threat assessment by mapping external intelligence and vulnerability information onto organizational, application, and technology layers of healthcare institutions, producing risk scores and actionable recommendations for security investments (Heinl et al., 2024). In both cases, AI supports healthcare processes by monitoring the infrastructure and data ecosystems on which clinical care depends.

Finally, process mining makes workflow analytics itself a HealthProcessAI application. The explicit HealthProcessAI framework wraps PM4PY and bupaR, uses LLMs to interpret discovered process maps, and generates structured reports for sepsis progression and chronic kidney disease pathways (Illueca-Fernandez et al., 29 Aug 2025). Here the object of analysis is the healthcare process as such.

5. Validation, deployment, and reported performance

Reported evaluations are heterogeneous because the systems differ in scope, modality, and maturity. Some are expert-reviewed prototypes; others are certified production systems; others are software frameworks validated through proof-of-concept case studies or deployment metrics. The table summarizes representative results already stated in the literature.

System	Process role	Reported result
Smart doctor app (Khan et al., 2018)	Symptom assessment and simple treatment recommendation	Working prototype covering more than 100 diseases
AITE at Medgate (Marchiori et al., 2020)	Remote medical triage	High-risk F-score ≈ 87.5%, medium-risk ≈ 74.0%, low-risk ≈ 90.4%
DeepSeek in tertiary hospitals (Chen et al., 23 Feb 2025)	Pathology, imaging, documentation, chronic care	Lung nodule differentiation accuracy > 95.2%; AI drafts ~80% of medical records
Multi-agent pre-consultation (Yu et al., 3 Nov 2025)	Triage and structured history taking	87.0% primary department triage; 98.2% task completion with agent-driven scheduling
Public-health monitoring system (Joshi et al., 4 Jun 2025)	National anomaly triage	54x increase in reviewer speed efficiency
MERLIN platform (Cohen et al., 2021)	Data acquisition and dataset generation	15,000 patient records per hour; cohort generation in 2–5 minutes

The smart-doctor prototype was tested multiple times and reviewed by professional medical doctors, who considered it accurate for its intended scope, although no confusion matrix or accuracy percentage was reported (Khan et al., 2018). AITE was trained on more than 900,000 teleconsultation case records and designed for transparency, user acceptance, and sub-1-second latency between questions; relation extraction also achieved strong scores, with the ELMo-Bi-LSTM model reporting $P=0.92$, $R=0.90$, and $F=0.91$ for the “not_located_in” relation (Marchiori et al., 2020).

Deployment-grade hospital systems show a different validation profile. DeepSeek is described as “certified and in use” and is evaluated through throughput, diagnostic accuracy, documentation efficiency, and operational gains rather than a single benchmark (Chen et al., 23 Feb 2025). The AI-driven Egyptian EHR reports ROUGE and BERTScore results for summary generation, with high recall and median BERT F1 greater than 0.85, while its ViT-based pneumonia model achieved loss 0.0868 and accuracy 0.9742 (Alorbany et al., 8 Feb 2025). The pre-consultation multi-agent system was evaluated on 1,372 validated EHRs and 18 physicians, yielding 80.5% secondary department classification, physician quality scores of 4.56 for Chief Complaints, 4.48 for History of Present Illness, and 4.69 for Past History, with consultations completed within 12.7 rounds for $T_2$ and 16.9 rounds for $T_3$ (Yu et al., 3 Nov 2025).

Framework papers emphasize operational and methodological validation. MERLIN demonstrates ingestion of 15,000 patient records per hour and collaborative cohort generation plus preliminary dataset analysis in 2–5 minutes (Cohen et al., 2021). The process-mining HealthProcessAI framework processed sepsis data across four proof-of-concept scenarios and compared five LLMs, with Claude Sonnet-4 and Gemini 2.5-Pro receiving evaluator means of 3.79/4.0 and 3.65/4.0 (Illueca-Fernandez et al., 29 Aug 2025). In public-health surveillance, the ranking-based system’s 54x gain in reviewer speed efficiency is a direct workflow metric rather than a classifier metric, and therefore especially characteristic of HealthProcessAI as a process-improvement paradigm (Joshi et al., 4 Jun 2025).

6. Governance, limitations, and future directions

The literature is consistent that HealthProcessAI is not synonymous with autonomous clinical replacement. Early systems that directly recommend medicines expose immediate safety issues: the smart-doctor paper itself implies risk of misdiagnosis outside its intended scope and raises concerns about recommending antibiotics such as Augmentin without prescription control (Khan et al., 2018). Production triage systems therefore emphasize conservative thresholds, red-flag logic, clinician override, and certification requirements (Marchiori et al., 2020).

Governance becomes more explicit as systems scale. DeepSeek’s rapid rollout in China’s tertiary hospitals is accompanied by a “responsibility gap,” concerns about automation bias, and proposals for shared accountability frameworks, real-time explainability, liability insurance pools, physician verification mandates, and institutional audit protocols (Chen et al., 23 Feb 2025). Model-driven and federated architectures answer a different governance problem: how to preserve semantic consistency and privacy across institutions. MILA, ontology validation, federated learning, and traceability ledgers are intended to make workflows reproducible, auditable, and compliant with GDPR, HIPAA, and related regulatory regimes (Raheem et al., 10 Oct 2025).

A persistent technical limitation is drift. The “Learning Machine” framework argues that healthcare models risk obsolescence as patient demographics, data-generating processes, and clinical practices change, and therefore must be continuously monitored through an online prediction loop and an offline safe-haven learning loop (Abroshan et al., 2023). This suggests that HealthProcessAI is best conceived as a monitored service with explicit retraining, recalibration, and audit policies rather than as a static deployed model.

Several forward paths recur across the corpus. The smart-doctor line proposes deeper trees, expansion from 3 to 7 levels, voice input, and more robust machine learning (Khan et al., 2018). Chronic-care and personalized-medicine systems point toward richer multimodal fusion of EHR, sensor, environmental, genomics, and radiomics data (Maranghi et al., 2022, Ivanovic et al., 2022). Process knowledge-infused AI argues for stronger guideline adherence and user-level explanations grounded in familiar concepts rather than opaque outputs (Sheth et al., 2022). The process-mining framework points toward broader live workflow analytics, predictive monitoring, and LLM-assisted translation of technical process models into clinician-facing narratives (Illueca-Fernandez et al., 29 Aug 2025).

Taken together, these directions indicate that HealthProcessAI is evolving toward a monitored, multimodal, privacy-aware, human-in-the-loop process layer for healthcare. This suggests a mature instantiation will combine explicit workflow models, heterogeneous data pipelines, decision support modules, and governance mechanisms so that AI can improve care pathways while remaining constrained by clinical oversight, validation, and institutional accountability.