The action of acupuncture can be explained by various levels of neuromodulations by means of local, segmental, and general effects (descending analgesia and central regulation) [1
]. Many neuroimaging studies have shown common brain activations and deactivation patterns following acupuncture stimulations [3
]. However, it has been difficult to reveal the point specificities on clinical significances using these approaches [4
]. Clinical acupoint selection involves three basic principles: (1) local acupoints near the area where symptoms occur, (2) distant acupoints from the symptom location along the meridian, and (3) distant acupoints based on symptom differentiation, regardless of symptom location [6
]. Recent studies have investigated these principles through the novel paradigm of data mining, using a number of available datasets related to traditional medicine. The associations between symptom differentiations and acupuncture prescriptions are relatively clear in classical textbooks, although the associations from these classical textbooks are not likely to be prominent in clinical practice [9
]. Previous studies have identified selections of acupoints for various diseases such as lumbar disc herniation, dysmenorrhea, rheumatoid arthritis, and visceral pain [11
]. However, these findings do not directly reveal the specificity of acupoint indication; they only indicate which acupoints can be selected specifically for a particular disease [15
In clinical practice, it is essential to identify which acupoints are specifically associated with a particular disease. However, selecting an acupoint to treat a particular disease does not always imply that the selected acupoint has specific indications for that disease [15
]. Thus, indications for the selected acupoints may not always match the target disease for which the acupoints were selected. This crucial point is often missed in both education and clinical practice, which often leads to errors in logic. To clarify this relationship, we suggested two-directional relationships between diseases and acupoints: forward inference and reverse inference. Forward inference analysis identifies the specificity of acupoint selection by analyzing the probability that an acupoint would be selected for a given disease: P (Acupoint | Disease). In contrast, reverse inference analysis identifies the specificity of acupoint indication by analyzing the probability of a given disease based on the selected acupoint: P (Disease | Acupoint). In our previous study, we proposed that the specificity of acupoint selection might differ from the specificity of acupoint indication. For example, BL23, GB30, and GV3 showed strong forward inferences to lumbar herniated intervertebral disc, while lumbar herniated intervertebral disc showed a high reverse inference score for acupoint BL23 [15
Implementation of forward and reverse inferences, based on a large database, is needed to clarify the associations between diseases and acupoints. The Cochrane Database of Systematic Reviews (CDSR) identifies, appraises, and synthesizes all empirical evidence that meets prespecified eligibility criteria for efficacy [16
]. Therefore, we hypothesized that we could identify the specificity of acupoint indication by analyzing relationships between prescribed acupoints and diseases using CDSR data. We also applied data mining techniques and Bayes factor (BF) correction to CDSR data. Data mining can reveal associations between diseases and acupoints [1
]. Bayes factor correction has been suggested as a way to avoid logical errors and fallacies in forward and reverse inference analyses [17
]. Bayes factor correction can help define the specificities of acupoint indications by correcting the odds of the prior probability or the odds of the diseases presented in the data. By eliminating the prior probability of diseases, the substance of reverse inference can be formalized to determine the specificities of acupoint indications.
In the current study, we searched acupuncture regimens in randomized controlled trials included in the CDSR and then identified the specificities of acupoint selections and acupoint indications. We explored the specificity of acupoint selections using forward inference and the specificity of acupoint indications using reverse inference based on clinical trial data.
The findings of the present study revealed patterns of acupoint selection for 30 diseases, as well as patterns of acupoint indications for 49 acupoints, using clinical trial data from the CDSR. We demonstrated the specificity of acupoint selection for each disease using forward inference. We also demonstrated the specificity of acupoint indication using reverse inference. Considering the prior probabilities of diseases, acupoint indications were further refined with BF correction. Finally, we identified two types of acupoint indications for 24 acupoints: regional and distal. Identification of the specific patterns of acupoint indications through data mining from clinical trials will be useful for understanding the characteristics of acupoints in education and clinical practice.
The core set of acupoints can be widely used to treat a variety of diseases, whereas some acupoints are only used to treat specific diseases. In our previous study, we identified the commonality and specificity of acupoint selections based on virtual acupuncture treatments prescribed by practicing clinicians. We found that acupoints ST36, LI4, and LR3 were the most commonly prescribed across all diseases [21
]. Mining of CDSR data also revealed that the main acupoints commonly used for a variety of pain management approaches were SP6, ST36, LI4, and LR3 [1
]. Consistent with the findings of previous studies, we found that the core acupoints were widely used for all 30 diseases, regardless of disease type. In contrast, acupoints GV20 and EX-HN3 were specifically used for the treatment of depression, while acupoints CV4 and SP8 were specifically used for the treatment of dysmenorrhea (Figure 2
). These findings support the common rule of acupoint selections, in which some acupoints are common among all diseases while other acupoints are specifically prescribed for the treatment of certain diseases.
Forward and reverse inference have been used in various areas of academic research. Notably, reverse inference has been used in neuroscience to assign a certain cognitive process to activation of a certain brain region, while forward inference has been used to evaluate cognitive theories based on different patterns of brain activation [18
]. As forward and reverse inferences can link associations directly, they demonstrate causal relationships without possible logical errors. This study extends the concept of forward and reverse inference in terms of the relationship between acupoints and diseases using CDSR data. The specificity of acupoint indication is generally inferred from the specificity of acupoint selection. However, we demonstrated that forward inference alone can lead to errors in logic regarding the acupoint indication specificity and that the application of reverse inference can improve the acupoint specificity.
Previous studies have indicated that reverse inference should be approached with caution if the prior probability (i.e., the odds of the prior probability of the disease) remains unknown [19
]. By implementing reverse inference, we found that depression and dysmenorrhea were specific indications for acupoint SP6, while depression, induction of labor, and subfertility were specific indications for acupoint LI4 (Figure 3
). However, the total number of trials can influence the specificity of acupoint indications. For example, depression was an indication for many acupoints (e.g., SP6, ST36, LR3, LI4) in the reverse inference because the total number of studies on depression was greater than the number of studies on other diseases. To avoid possible fallacy following the distortion of data, BF correction was implemented when identifying the specificity of acupoint indications; this approach corrected the odds of the prior probability of the disease. Implementation of BF correction led to the identification of insomnia as the indication for acupoint HT7 while ruling out diseases with high frequency in the data (e.g., depression) and specific acupoints (e.g., EX-HN3) (Figure 4
). This analysis demonstrated that reverse inference using BF correction is useful for inferring specificity of acupoint indications.
In the current study, two types of acupoint indications were identified: regional and distal. Both regional and distal acupoints were specifically linked with certain diseases. Among regional acupoints, low back pain was the indication for acupoint BL23 in the back, while acute ankle sprain was the indication for acupoint BL60 in the feet. Among distal acupoints, low back pain was the indication for BL40 in the lower limb, while insomnia was the indication for acupoint HT7 in the wrist (Figure 5
). Classification of acupoint indications into regional and distal types is consistent with the approach used in previous studies, where both regional and distal acupoints were mainly used for pain management during acupuncture, based on data mining from clinical trials [1
Two different strategies for acupoint prescription, including branch treatment and root treatment, are widely used in the context of traditional East Asian medicine [22
]. Pattern identification can extract and synthesize patients’ signs and symptoms and lead to making treatment decisions [23
]. Acupuncture practitioners can choose the most appropriate acupoints based on the disease or symptoms (i.e., branch treatment) or on the results of pattern identification (i.e., root treatment) [21
]. Although some clinical trials consider individualized treatment strategies, clinical trials are generally conducted to find out the efficacy of the acupuncture treatment on a certain disease. Furthermore, pattern identification can be varied across patients even though they have the same disease [25
]. Only a limited number of studies provided the information of pattern identification from the CDSR. Thus, the present study only determined the acupoint indication based on the relationship between disease/symptoms and acupoints without considering pattern identification. The acupoint indication associated with pattern identification should be further studied in the future.
Our study had several limitations. First, it did not consider the effectiveness of the acupuncture treatment in each study. To include the effectiveness of acupoints, future studies with appropriate control groups will be needed to enable the extraction of the effectiveness of acupuncture treatment administered to a group of acupoints. Due to the limited number of included studies, we were not able to analyze the data considering the clinical effectiveness of needling at each point. More sufficient data will be needed to ensure the acupoint indications based on clinical significance in the future. Second, this study did not restrict methodological quality among the assessed studies. Insufficient evidence was available to reveal discrepancies between acupoints used in high-quality and low-quality studies. Future analyses should incorporate inclusion criteria to ensure the inclusion of studies with meaningful information. Finally, this study identified acupoint indications only from clinical trials. Additional analyses will be needed to clarify the associations between diseases and acupoints based on clinical data obtained in a real-world setting.
In summary, data mining, forward and reverse inference analyses, and BF correction of clinical trial data revealed the bidirectional specificity of acupoints for various diseases. We expect our approach to provide new information regarding the specificities of acupoint indications from clinical observations.