Changes in Evidence for Hemiplegia and Upper Limb Paralysis Training in Stroke Treatment Guidelines from the 2004 to 2023 Editions
The approach to hemiplegia and upper limb paralysis after stroke in the first edition of the 2004 Stroke Treatment Guidelines focused primarily on spasticity management and basic functional maintenance, based on the evidence available at that time.
Compared to the 2009 and 2023 editions, this guideline had limited treatment options and scientific evidence, centering mainly on pharmacological interventions and basic rehabilitation.
Below, the content of the 2004 edition is described in detail, supplemented with the latest information and research trends as of 2025.
In the 2004 edition, the management of spasticity associated with hemiplegia was a major pillar of treatment.
For spasticity, the use of muscle relaxants such as dantrolene sodium, tizanidine, baclofen, and diazepam was recommended.
These drugs aimed to maintain patients' activities of daily living (ADL) by suppressing muscle tone, thereby reducing joint range of motion limitations and pain caused by spasticity.
In particular, for baclofen, in addition to oral administration, intrathecal therapy was considered for severe spasticity, with sustained effects expected.
As for non-pharmacological interventions, stretching and range-of-motion exercises were recommended and considered to contribute to the prevention and reduction of spasticity.
Transcutaneous electrical nerve stimulation (TENS) was proposed as an aid for reducing spasticity and maintaining muscle strength, but the evidence was insufficient.
Functional electrical stimulation (FES) was also suggested to be potentially useful for maintaining muscle strength and movement patterns, but no standardized protocol had been established.
Botulinum toxin therapy was mentioned in some parts as being effective for reducing spasticity, but as of 2004, it was not covered by insurance in Japan, and its clinical use was limited.
Nerve blocks using phenol or ethyl alcohol were considered effective for improving joint range of motion limitations caused by spasticity, but due to technical difficulties and the risk of side effects, they were performed only in specialized facilities.
For upper limb paralysis, the use of orthoses was recommended, and in particular, orthoses that keep the paralyzed upper limb in an extended position were considered useful for spasticity management. Cooling and heat therapy were considered as auxiliary means, but scientific evidence was lacking.
Regarding functional recovery of upper limb paralysis, the 2004 edition had limited descriptions of specific rehabilitation methods.
Basic training through physical therapy and occupational therapy was the focus, and exercises utilizing the patient's residual functions were recommended.
However, specific methods such as task-oriented training or constraint-induced movement therapy (CIMT), which are emphasized in later guidelines, were hardly mentioned.
This reflects the fact that rehabilitation research at the time had not yet accumulated sufficient evidence regarding neuroplasticity and the mechanisms of functional recovery.
Electrical stimulation therapy (FES and TENS) was proposed as an aid for maintaining muscle strength and motor relearning, but there were large individual differences in effectiveness, and clinical standardization had not progressed.
Based on the latest information and research trends as of 2025, the 2004 edition of the guidelines is limited from a modern perspective, but it was a realistic guideline in the medical environment of that time.
Botulinum toxin therapy was not covered by insurance in 2004, but since 2021, insurance coverage for spasticity in the upper and lower limbs has expanded in Japan, and it has become widely used as a standard treatment.
Recent studies have examined the timing of botulinum toxin administration (early post-onset vs. chronic phase), optimization of dosage, and the effects of simultaneous administration to multiple muscle groups, emphasizing the importance of individualized treatment plans for each patient.
Regarding FES and TENS, technological advancements have enabled precise adjustment of stimulation parameters, leading to an increase in evidence contributing to muscle strength recovery and improved motor control.
Robot-assisted rehabilitation, which was barely mentioned in 2004, has seen significant development in AI-powered rehabilitation equipment since the 2020s, enabling improved precision in repetitive movements and individualized training.
On social media, patients and healthcare professionals are discussing the effects of robot therapy and non-invasive brain stimulation (repetitive transcranial magnetic stimulation: rTMS, transcranial direct current stimulation: tDCS), though some note that verification through large-scale randomized controlled trials (RCTs) is still lacking.
Non-invasive brain stimulation was not addressed at all in the 2004 edition, but research in the 2020s has shown its potential to promote neuroplasticity, with effectiveness reported in patients with mild to moderate paralysis or in the early post-onset phase.
Visual stimulation (e.g., mirror therapy) and training using virtual reality also did not exist in 2004, but in recent years, they have gained attention for improving patient motivation and as integrated rehabilitation combined with cognitive function.
However, the accumulation of evidence for these approaches is still ongoing, and further research is required for standardization.
Discussion
Looking back at the 2004 edition of the Stroke Treatment Guidelines, treatment for hemiplegia and upper limb paralysis was strongly focused on the management of spasticity.
This indicates that spasticity was a priority issue to be addressed in clinical practice, as it was a major problem causing joint contractures, pain, and limitations in ADLs.
The recommendation of pharmacological interventions (baclofen, tizanidine, etc.) and nerve blocks aimed at the physiological control of spasticity, reflecting the era's background where symptom relief was prioritized over functional recovery.
The reason descriptions regarding functional recovery of upper limb paralysis were limited is that rehabilitation research at the time had not yet reached an understanding of neuroplasticity or the establishment of specific functional recovery methods (e.g., CIMT, task-oriented training).
This point is a limitation of the 2004 edition, while simultaneously clearly demonstrating the constraints of scientific knowledge at that time.
Comparing the evolution from the 2004 edition to the 2009 and 2021 editions, the guidelines have shifted significantly from spasticity management to a focus on functional recovery alongside the accumulation of evidence.
In the 2009 edition, the effects of botulinum toxin and electrical stimulation were shown more clearly, and in the 2021 edition, CIMT and robot therapy became strongly recommended.
This change is supported by basic research (elucidation of neuroplasticity), clinical research (increase in RCTs), and technological advancements (robotics, AI, and brain stimulation technology).
While the conservative approach of the 2004 edition may seem insufficient from a modern perspective, it can be said that it was a realistic guideline that contributed to maintaining patients' QOL within the scope of medical resources and knowledge available at the time.
As a discussion based on the latest research as of 2025, the possibilities and challenges of technology-driven rehabilitation are highlighted.
Robot-assisted therapy and AI-powered rehabilitation equipment are excellent in providing precision in repetitive movements and individualized training, but high costs and barriers to introduction in medical institutions remain challenges.
Non-invasive brain stimulation (rTMS, tDCS) is a promising method for promoting neuroplasticity, but individual differences in effectiveness and the establishment of optimal protocols remain unresolved.
While there is high anticipation for new technologies in discussions on social media, there are also opinions that further verification is necessary for practical application in clinical settings.
Future challenges include establishing personalized treatments based on patient conditions (onset time, severity of paralysis, comorbidities) and improving access to treatment through education for healthcare professionals and patients.
Creating an environment where patients can understand their treatment options and actively participate in rehabilitation will be key to functional recovery and improving quality of life.
Treatment approaches for post-stroke hemiplegia and upper limb paralysis in stroke treatment guidelines from the 2004 edition to the 2023 edition (the latest as of 2025) have evolved alongside the accumulation of scientific evidence.
The following describes the details, focusing on the 2009 and 2021 editions and including the latest research and information. Note that since public information on guidelines from the 2023 edition onwards is limited, this description is based on the 2021 edition while supplementing with the latest research trends.
In the 2009 edition of the guidelines, the management of spasticity associated with hemiplegia was a primary focus. Pharmacotherapy such as dantrolene sodium, tizanidine, baclofen, diazepam, and tolperisone was recommended for spasticity. These drugs are used to reduce spasticity by suppressing muscle tone, with the goal of improving the patient's functional prognosis. In particular, tizanidine was evaluated as having efficacy equivalent to baclofen and diazepam but with fewer side effects. For significant spasticity, intrathecal baclofen therapy was effective, and long-term sustained effects were confirmed. Additionally, for limitations in joint range of motion, nerve blocks using phenol or ethyl alcohol were recommended, and improvements in the modified Ashworth scale and joint range of motion were reported. Botulinum toxin (which was not covered by insurance at the time) was considered effective for reducing spasticity in the upper and lower limbs and improving activities of daily living (ADL), with injections into the upper arm, forearm, and finger muscle groups being particularly effective. Furthermore, as non-pharmacological interventions, transcutaneous electrical nerve stimulation (TENS), stretching, range of motion training, and the use of orthoses with functional electrical stimulation (FES) were recommended. However, cooling and heat therapy were considered to have insufficient scientific evidence.
In the 2021 edition of the guidelines, the approach to hemiplegia and upper limb paralysis shifted to focus more on functional recovery. For mild to moderate upper limb paralysis, task-oriented training and training that forces the use of the paralyzed upper limb (e.g., constraint-induced movement therapy, CIMT) are strongly recommended. These trainings have been confirmed to improve motor function by repeating specific movements, and in particular, to enhance practicality in daily life. CIMT is a method that promotes functional recovery by suppressing the use of the non-paralyzed upper limb and actively using the paralyzed one, and it can be widely introduced as it does not require special equipment. However, there are reports that its effectiveness is limited in the period before the subacute phase. Robot-assisted upper limb functional training is also considered effective, and its utility is recognized, particularly in accurately supporting repetitive movements. For moderate to severe upper limb paralysis and shoulder subluxation, neuromuscular electrical stimulation (NMES) is recommended, contributing to improvements in muscle strength and joint stability. Training utilizing visual stimulation (motor observation or mirror therapy) and motor imagery is also considered effective, and approaches using virtual reality are attracting attention. Repetitive transcranial magnetic stimulation (rTMS) and transcranial direct current stimulation (tDCS) are considered after taking patient selection and safety into account, but additional effects when combined with task-oriented training may not be clear. Other methods, such as peripheral vibration stimulation, sensory stimulation, and vagus nerve electrical stimulation, are also being attempted, but sufficient evidence is still lacking.
Regarding the guidelines from the 2023 edition onwards, although public information is limited, several advancements can be inferred from recent research trends. First, botulinum toxin therapy has seen an expansion in insurance coverage for some indications in Japan since 2021 and is becoming more widespread as a standard treatment in spasticity management. Recent research is examining the optimization of administration timing and dosage of botulinum toxin, as well as the effects of simultaneous administration to multiple muscle groups, emphasizing the importance of personalized treatment protocols. Also, with the evolution of robotic technology, the precision and effectiveness of robot-assisted therapy in upper limb rehabilitation are improving. In particular, AI-utilized rehabilitation equipment is moving toward analyzing patient movement patterns in real-time and providing customized training. Non-invasive brain stimulation such as rTMS and tDCS continues to be researched for its potential to promote post-stroke neuroplasticity, and effectiveness in specific patient groups (e.g., patients early after onset) has been reported. On the other hand, training using sensory stimulation and virtual reality is attracting attention as an approach to improve patient motivation and integrate with cognitive function, but verification through large-scale randomized controlled trials (RCTs) is required. Information on social media shows patients and healthcare professionals discussing the introduction and effects of these new technologies, but further data is needed to establish scientific evidence.
Discussion
Considering the transition of the guidelines from the 2009 edition to the 2021 edition, it is clear that treatment for post-stroke hemiplegia and upper limb paralysis has shifted its focus from spasticity management to functional recovery. In the 2009 edition, the reduction of spasticity was the primary goal, and pharmacological and non-pharmacological interventions were central. This reflects a period when suppressing muscle tone was prioritized because spasticity has a direct negative impact on joint range of motion and ADL. In particular, the recommendation of botulinum toxin and nerve blocks was an approach focused on the physiological control of spasticity. On the other hand, in the 2021 edition, active rehabilitation aimed at functional recovery is emphasized, and methods that encourage active patient participation, such as task-oriented training and CIMT, have become mainstream. This change is considered to be the result of progress in the understanding of neuroplasticity and the recognition of the importance of patient-centered approaches in rehabilitation.
Based on the latest research trends, the utilization of technology has become a new frontier in treatment.
Robot-assisted therapy and AI-utilized rehabilitation are promising in that they increase the precision of repetitive movements and enable training that meets the individual needs of patients.
However, these technologies are high-cost, and accessibility for medical institutions and patients becomes a challenge.
Also, brain stimulation technologies such as rTMS and tDCS are expected to be treatments based on neuroscientific evidence, but individual differences in effectiveness and the establishment of optimal protocols remain unresolved. In discussions on social media, while there is high anticipation for these new technologies, there are also voices saying that time and verification are necessary for actual introduction in clinical settings.
Future challenges include the further accumulation of evidence and the promotion of personalized medicine.
In particular, the construction of optimal treatment protocols according to the patient's condition (onset time, severity of paralysis, comorbidities, etc.) is required.
Also, as the guidelines evolve, education for healthcare professionals and the provision of information to patients will become important.
Creating an environment where patients can understand their own treatment options and actively participate in rehabilitation will be the key to functional recovery and improved quality of life.
Ranking: Utilization of Generative AI in Improving Robot-Assisted Therapy and AI-Enhanced Rehabilitation
1st: ChatGPT (OpenAI)
2nd: Gemini (Google)
3rd: Claude (Anthropic)
4th: Copilot (Microsoft)
5th: DeepSeek (DeepSeek AI)
6th: Grok (xAI)
7th: Perplexity AI
8th: Llama 3 (Meta AI)
9th: Grok 3 mini (xAI)
10th: Genspark
11th: Claude 3.5 Sonnet (Anthropic)
12th: ChatGPT o3 (OpenAI)
13th: Gemini 1.5 Pro (Google)
14th: DeepSeek R1 (DeepSeek AI)
15th: Microsoft 365 Copilot (Microsoft)
16th: Claude 3.7 Sonnet (Anthropic)
17th: Gemini 2.0 Flash (Google)
18th: Grok 2 (xAI)
19th: ChatGPT Pro (OpenAI)
20th: Claude 3 Haiku (Anthropic)
Detailed information, latest updates, and recent research
ChatGPT (OpenAI) ranks at the top for improving robot-assisted therapy and AI rehabilitation, leveraging its high natural language processing capabilities and diverse data processing abilities among generative AIs.
Its Deep Research feature is useful for rapidly analyzing vast amounts of medical literature and rehabilitation research, proposing personalized training programs, and optimizing robot control algorithms.
According to the latest 2025 information, the GPT-4o model of ChatGPT has enhanced real-time data processing and multimodal support (text, image, audio), improving its ability to analyze patient movement data and dynamically adjust rehabilitation plans.
In particular, when generating task-oriented training protocols for post-stroke upper limb paralysis, it is capable of making specific suggestions based on the patient's movement patterns and recovery stage.
On social media, there have been reports of healthcare professionals using ChatGPT to efficiently create rehabilitation guidelines, though verification of accuracy is required for tasks requiring specialized knowledge.
Gemini (Google) has a strength in its integration with Google's search engine and cloud services, excelling in the ability to collect and analyze the latest medical research and clinical data in real time.
Gemini Deep Research is suitable for data analysis in robot-assisted stroke rehabilitation, processing patient movement logs and sensor data to propose personalized training parameters.
In 2025 research, it was reported that Gemini 2.0 has enhanced multimodal processing, for example, by analyzing patient movement videos to generate algorithms that optimize robot movements.
Integration with Google Workspace also streamlines data sharing and report creation in medical institutions.
However, there are remaining challenges in local support, such as instances where optimization for Japanese medical literature is insufficient and pricing information is displayed in dollars.
Claude (Anthropic) is characterized by a design that emphasizes safety and ethics, making it suitable for accurate information processing in the medical field.
Claude 3.5 Sonnet excels in long-form processing and natural Japanese generation, and is utilized for designing rehabilitation protocols and creating patient education materials.
In the latest 2025 trends, Claude Advanced Research is being evaluated for its ability to integrate internal medical institution documents with external data to generate rehabilitation plans tailored to the individual needs of stroke patients.
On social media, there are comments that Claude's reports are easy to read and practical for medical professionals, but its processing speed can sometimes be slower compared to ChatGPT or Gemini.
Copilot (Microsoft) has high compatibility with existing medical institution systems due to its integration with Microsoft 365. In the 2025 update, Copilot's 'Think Deeper' feature is being utilized for data analysis in rehabilitation research, enabling the visualization and automation of training plans using Excel and PowerPoint.
In stroke rehabilitation, cases have been reported where patient exercise data is processed in Excel to build feedback loops for robot-assisted therapy.
However, the Deep Research feature is limited to the Enterprise plan, posing challenges for accessibility for individuals and small-to-medium-sized medical institutions.
DeepSeek (DeepSeek AI) is gaining attention as a cost-effective option, specializing in mathematics and scientific data analysis.
In 2025 research, DeepSeek R1 was used to optimize control algorithms for robot-assisted therapy, contributing to improved accuracy in repetitive movements.
However, there are challenges with the naturalness of its Japanese processing, and its application in the medical field is mainly limited to data analysis.
On social media, while its ability to perform high-performance analysis for free is appreciated, concerns regarding data privacy for Chinese-made AI have been pointed out.
Grok (xAI) has a strength in real-time data access, making it suitable for grasping the latest trends in rehabilitation research by utilizing information from social media (X).
Grok 3 was released in 2025, with improved reasoning capabilities in mathematical and scientific tasks, which is useful for motion analysis in robot-assisted therapy.
However, there are issues with the accuracy of the DeepSearch feature, and it is considered necessary to use it in conjunction with other AIs for important decision-making.
On social media, there are opinions that Grok's humorous responses help improve patient motivation for education, but further improvements are needed to enhance its reliability in the medical field.
Perplexity AI, as an information-retrieval AI, excels in its ability to quickly collect the latest medical literature and clinical data.
In 2025 trends, it is being utilized to generate evidence-based protocols for stroke rehabilitation, but it can sometimes lack depth in analysis.
Llama 3 (Meta AI) is open-source, making it highly customizable and used in research institutions for developing robot control algorithms, but there are restrictions on commercial use.
Grok 3 mini and Genspark are characterized by lightweight and fast processing, but they are unsuitable for generating complex rehabilitation plans.
Claude 3.7 Sonnet, ChatGPT o3, and Gemini 1.5 Pro are high-performance as the latest models for each, but they are positioned in the middle of the rankings because their degree of specialization for specific rehabilitation tasks varies.
Regarding accessibility challenges, ChatGPT and Gemini offer free plans or low-cost trials, lowering the barrier to entry for medical institutions and patients.
Claude and Copilot are primarily paid plans, but their strength lies in integration features for businesses. DeepSeek and Grok can be used for free, but challenges regarding data privacy and Japanese language support remain.
On social media, attempts by medical professionals to combine these AIs to balance cost and effectiveness have been reported.
Discussion
Utilizing generative AI (such as ChatGPT, Gemini, Claude, Copilot, DeepSeek, and Grok) to improve robot-assisted therapy and AI rehabilitation holds the potential to open new frontiers in the treatment of post-stroke hemiplegia and upper limb paralysis.
ChatGPT, ranked first, is capable of handling a wide range of tasks, from analyzing patient data to generating personalized training programs, thanks to its high versatility and multimodal support.
Its Deep Research feature can rapidly process medical literature and clinical data, contributing directly to motion control in robot-assisted therapy and the optimization of rehabilitation plans.
On the other hand, the high cost of the Pro plan ($200/month) can be a barrier, leaving challenges regarding accessibility for small-to-medium-sized medical institutions and individual patients.
Gemini's strength lies in its real-time data collection capability through integration with the Google ecosystem, making it suitable for proposing rehabilitation strategies that reflect the latest research trends.
However, its insufficient optimization for Japanese medical literature may limit its use in local clinical settings.
Claude is designed with ethics and safety in mind, which enhances its reliability in the medical field and makes it excellent for creating patient education materials and documenting rehabilitation plans, but its slower processing speed affects its immediacy in clinical practice.
Copilot, through its integration with Microsoft 365, has high compatibility with existing medical institution systems and contributes to the efficiency of data-driven rehabilitation, but reliance on the Enterprise plan poses a cost barrier.
DeepSeek offers high-performance data analysis at a low cost, but concerns regarding Japanese language support and data privacy reduce its reliability in the medical field.
Grok's strength is its real-time data access, but due to accuracy issues, its role remains largely that of an auxiliary tool.
Based on the latest research trends, the use of generative AI is bringing significant progress in improving the precision of repetitive movements in robot-assisted therapy and in generating training programs tailored to individual patient needs.
For example, motion analysis using ChatGPT or Gemini can process sensor and video data in real-time to optimize robot feedback loops.
However, the introduction of these technologies is hindered by high costs, the need for specialized knowledge, and data privacy issues.
On social media, there are attempts by healthcare professionals to use combinations of generative AI to balance cost and effectiveness, but since verification through large-scale RCTs is lacking, establishing evidence remains a future challenge.
Deepening the discussion, while the use of generative AI promotes the personalization and efficiency of stroke rehabilitation, it also carries the risk of widening the gap in accessibility.
High-performance models like ChatGPT and Gemini are being adopted by large medical institutions and research facilities, but costs remain a barrier for small local facilities and low-income patients.
The free plans of DeepSeek and Grok may mitigate this gap, but unless issues regarding reliability and Japanese language support are resolved, full-scale adoption in the medical field will be difficult.
Furthermore, the risk of AI hallucinations (generation of misinformation) is particularly critical in the medical field, making a verification process for generated rehabilitation plans and data analysis results essential.
Future guidelines will likely require ethical guidelines for the medical application of generative AI and cost-reduction measures (e.g., the use of open-source models).
To achieve patient-centered care, in parallel with advancements in AI technology, it is necessary to provide education for healthcare professionals and patients, as well as policy support to increase accessibility.