A machine learning-based hybrid recommender framework for smart medical systems

Encoding module

The index data encoding module 120 comprises two main units: the embedding vectorization unit 121 and the context encoding unit 122, respectively. The embedding vectorization unit 121 converts multiple index data into an embedding vector sequence by employing the embedding layer of the context encoder. The context encoding unit 122 implements the converter-based Bert model to perform semantic encoding on global-based context by using the sequence of embedding vectors, generating multiple semantic feature vectors that capture the correlation between different index data items in each department (Liu & Zhao, 2023). The proposed index data coding module performs contextual semantic encoding on index data by implementing DLMs to extract high-dimensional hidden features associated with the correlation between different index data items.

The multi-scale feature extraction module 130 concatenates multiple index semantic feature vectors into department feature vectors and passes them through the multi-scale neighborhood feature extraction module. So, multi-scale department feature vectors are attained corresponding to each department. Then, CNN is implemented for local feature extraction, where the convolution kernel moves along the time dimension in the form of a sliding window and outputs a weighted sum of the data within each time-series segment (Alpaslan & Hanbay, 2020). Large and small-scale convolution kernels are combined to extract features of different timing scales, achieving smooth input data, and minimizing information loss.

Suppose that the size of the input data is denoted by (n, t, d), where n, t, and d represent the sample size, time series length, and the characteristic dimension of each time step, respectively. In CNN, k convolution nuclei of different sizes are usually employed to extract multi-scale neighborhood features, where the size of the ith convolution kernel is denoted by (h_i, d, c_i), where h_i and c_i represent the window size, and the number of the convolution kernel, respectively. Then, for a convolution kernel, its weight, bias, and output vector can be defined as ${\mathrm{W}}^{(\mathrm{i})}={\mathrm{R}}^{\mathrm{h}\mathrm{i}\cdot \mathrm{d}\cdot \mathrm{c}\mathrm{i}},{\mathrm{b}}^{(\mathrm{i})}={\mathrm{R}}^{\mathrm{c}\mathrm{i}},\mathrm{a}\mathrm{n}\mathrm{d}{\mathrm{Z}}^{(\mathrm{i})}={\mathrm{R}}^{\mathrm{n}(\mathrm{t}-\mathrm{h}\mathrm{i}+1)\mathrm{c}\mathrm{i}}$, respectively. The convolution operation is conducted by the convolution checking on the input data. Equation (1) represents this relation.

(1) $$\mathrm{Z}=\mathrm{f}({\sum }_{j=1}^d{w}_j^{\left(i\right)}\ast x_j+b^{\left(i\right)})$$where * and f stand for convolution operation and activation function, respectively. The computed output of the convolution kernel is a tensor represented by (n, t − h + 1, c_i) magnitude, where each time step corresponds to a c_i dimensional eigenvector. Often, pooling is also required to reduce feature dimensions and model complexity. The multiple convolution kernels of output can be pieced together, forming a (n, t − h + 1, $\sum _{i=1}^k{c}_i$) size of a tensor, among them, the $h=\mathrm{m}\mathrm{a}\mathrm{x}\{{ℎ}_1,\dots ,{ℎ}_k\}$ denotes the maximum window size in the CNN model. This tensor is the multi-scale neighborhood feature, which can be directly employed in the subsequent classification task.

To sum up, mathematical expressions are required to describe the CNN model and the specific implementation process of its convolution operation, including relevant parameters such as weight matrix, bias vector, and activation function, and multi-scale neighborhood features extracted through convolution operation and pooling operation, respectively.

Feature extraction module

Figure 4 shows the block diagram of the multi-scale feature extraction module 130 in the registration review system of the medical institution.

The module comprises several units: concatenation unit 131, first neighborhood scale encoding unit 132, second neighborhood scale encoding unit 133, and multi-scale feature cascading unit 134. Concatenation unit 131 combines multiple index semantic feature vectors to obtain department feature vectors. First neighborhood scale encoding unit 132 applies one-dimensional convolution processing with a kernel length on the department feature vector to derive the first scale department feature vector. Second neighborhood scale encoding unit 133 applies one-dimensional convolution processing with a different kernel length on the department feature vector to generate the second scale department feature vector. Lastly, multi-scale feature cascading unit 134 concatenates the first and second-scale department feature vectors to produce multi-scale department feature vectors for each department. These feature vectors capture the relationship between different index data at different scales.

The eigenvector correction module 140 (ECM) performs eigenvalue corrections based on multi-scale department feature vectors, generating corrected vectors for a better representation of medical institution service levels. Thus, corrected vectors are arranged in a two-dimensional feature matrix and utilized as a feature extractor for a CNN model. However, coding fluctuations of index-related semantic features at multiple scales may cause phase differences between corresponding department feature vectors. Such differences can affect the classification accuracy of the feature map by hindering proper aggregation of the matrix. To address this issue, the system aligns multi-scale department feature vectors by implementing advanced techniques before arranging them into a feature matrix. With aligned vectors, related features occupy consistent positions within the matrix, improving the aggregation effect on the class probability of the classifier and boosting classification accuracy.

Specifically, the eigenvector correction module 140 is further configured to: employ Eq. (2) to perform eigenvalue correction based on the maximum value on the multi-scale department eigenvectors that correspond to each department to obtain the corrected multi-scale department features relevant to each department vector. Equation (2) gives this vector.

(2) $${\mathit{V}}^{\mathbf{\prime }}=\mathit{V}\odot {\mathit{e}}^{-\mathrm{s}\mathrm{i}\mathrm{n}\left(2\pi \odot \mathit{V}\odot {v_{max}}^{-1}\right)}$$where ${\mathit{V}}^{\mathbf{\prime }}$, $\mathit{V}$, and $v_{max}$ represent the multi-scale department feature vector, the corrected multi-scale department feature vector, and the maximum eigenvalue of the multi-scale department feature vector, respectively.

The proposed method employs the aggregation of the wave function representation to address the challenge of phase differences between multi-scale department feature vectors. It employs amplitude to represent intensity information and phase knowledge to characterize periodic position information, enabling complex vector information representations. This mitigates negative impacts on class probability aggregation, leveraging in-phase enhancement and out-of-phase cancellation based on the wave function to enhance classification accuracy. The approach improves the rationality and accuracy of the service level evaluations of medical institutions. Overall, the system leverages the DLMS and the aggregation of the wave function representation for more accurate and reliable service-level evaluations of medical institutions.

Associative encoding module

The research employs the inter-department association encoding module 150 to arrange department feature vectors into a 2D matrix. The CNN model is then employed to extract classification features. Correlation characteristics are extracted within each department. To consider the linkage relationships between departments, the corrected multi-scale department feature vectors are arranged into a 2D matrix and CNN is implemented to attain classification feature maps, extracting associated features between departments.

The inter-department association encoding module 150 is also employed for each CNN layer to perform the forward passes of input data. A convolution unit performs 2D convolution processing by implementing the convolution kernel. Then, a pooling unit performs mean pooling on the convolutional feature map based on the local feature matrix. An activation unit performs nonlinear activation on the pooled feature map to obtain the activation feature map. The output of the last CNN layer is a classification feature map.

The whole process can be expressed by Eq. (3).

(3) ${\mathrm{F}}_{\mathrm{o}\mathrm{u}\mathrm{t}}=\mathrm{C}\mathrm{N}\mathrm{N}({\mathrm{F}}_{\mathrm{i}\mathrm{n}})$where F_in denotes the input data, CNN represents the convolutional neural network model, and F_out characterizes the output feature map. In this process, the CNN model converts the two-dimensional feature matrix into the classification feature map through multi-layer convolution, pooling, and activation operations, respectively, to realize the classification diagnosis of patients.

Equation (4) represents the specific formula for each step in the inter-department association coding module 150.

(1) For each department, a multi-scale feature extractor was utilized to encode its indicator data and extract the correlation features among indicators. The feature graph $\mathrm{X}\in {\mathrm{R}}^{\mathrm{H}\times \mathrm{W}\times \mathrm{C}\mathrm{i}\mathrm{n}}$ was input into a multi-scale feature extractor to obtain feature graph ${\left\{{\mathrm{X}}^{\mathrm{s}}\right\}}_{\text{s}=1}^{\text{s}}$ with different scales, where S represents the number of scales. Equation (4) presents the feature map of each scale.

(4) ${\mathrm{X}}^{\mathrm{s}}={\mathrm{f}}_{\mathrm{s}}(\mathrm{X}),\mathrm{s}=1,2,\dots ,\mathrm{S}$where f_s(·) represents the eigenfunction employed to derive the s scale.

(2) The corrected multi-scale department feature vectors corresponding to each department were arranged into a two-dimensional feature matrix, which contained the correlation features among the internal indicators of each department and the indicators of each department. For each department, its corresponding multi-scale feature vectors ${\left\{\mathrm{X}\mathrm{k}\right\}}_{\text{k=1}}^{\text{ki}}$ were arranged into a two-dimensional feature matrix according to certain rules ${\mathrm{M}}_{\mathrm{i}}\in {\mathrm{R}}^{\mathrm{H}\times \mathrm{W}\times \mathrm{C}}$, where K_i represents the number of samples in the i-th department, C = C_in × S represents the length of feature vector at each position. Equation (5) presents the process.

(5) $${\mathrm{M}}_{\mathrm{i}}=(\mathrm{h},\mathrm{w},\mathrm{c})=\{\begin{array}{c}{x}_k^s\left(h^{\mathrm{\prime }},w^{\mathrm{\prime }},c^{\mathrm{\prime }}\right),if\left(h^{\mathrm{\prime }},w^{\mathrm{\prime }},c^{\mathrm{\prime }}\right)\in P_{\left(h,w,c\right)}\\ 0,otherwise\end{array}$$

(h′, w′, s′, c′) is an element of the eigenvector X^s_k and P_(h,w,c) is a set of all of the elements that need to extract (h′, w′, s′, c′) collection.

(3) The CNN model was employed to process the input data and obtain the output feature map. The two-dimensional feature matrix M_i was input into the CNN and the classified feature graph ${\mathrm{F}}_{\mathrm{i}}\in {\mathrm{R}}^{\mathrm{H}\mathrm{o}\times \mathrm{W}\mathrm{o}\times \mathrm{C}\mathrm{o}}$ was obtained, where H_o, W_o, and C_o represent the height, width, and number of channels of the output feature graph, respectively. Equation (6) represents the CNN.

(6) ${\mathrm{F}}_{\mathrm{i}}=\mathrm{C}\mathrm{N}\mathrm{N}({\mathrm{M}}_{\mathrm{i}})$where the CNN(·) represents the convolutional neural network model, and F_i denotes the output feature map.

(4) The output of the last layer was the classification feature map, which was implemented to classify patients. The classification feature graph {F_i}^N_{i = 1} of all departments was summarized to obtain the final classification feature graph F_out ∈ R^Ho×Wo×N, where N represents the number of departments. Equation (7) represents this process.

(7) ${\mathrm{F}}_{\mathrm{o}\mathrm{u}\mathrm{t}}(\mathrm{h},\mathrm{w},\mathrm{i})={\mathrm{F}}_{\mathrm{i}}(\mathrm{h},\mathrm{w}),\mathrm{i}=1,2,\dots ,\mathrm{N}$where (h,w) denotes a position in the output feature graph. Finally, the classification algorithm can be employed to classify the classification feature map, to realize the classification diagnosis of patients.

Classification module

The result generation module of reviews (160) is employed to pass the classification feature map through a classifier to obtain a classification result that is utilized to represent the service level label of the medical institution to be reviewed.

The result generation module of evaluations (160) is further employed to process the classification feature vector presented in Eq. (8) to obtain the classification result by employing the classification feature map.

(8) $$O=f\left(W,b\right)={\left[\sum _{i=1}^{k}exp\left(W_{i}x+b_i\right)\right]}^{-1}\left[\begin{array}{c}exp\left(W_{1}x+b_1\right)\\ exp\left(W_{2}x+b_2\right)\\ .\\ .\\ .\\ exp\left(W_{k}x+b_k\right)\end{array}\right]$$where O, W_i, and b_i denote the output matrix, and the weight and bias matrices corresponding to the ith classification, respectively.

Other functional modules

The registration review system for a medical institution has several functional modules, including a login module, a SaaS background management module, an ICD-10 dictionary unit, an ICD-9-cm-dictionary unit, an M-code dictionary unit, an implementation plan module, and a data collection module. For example, the SaaS background management module is further divided into an organization management unit, a personnel management unit, an authority management unit, an ICD-10 dictionary unit, an ICD-9-cm-dictionary unit, and an M code Dictionary unit, a review standard unit. These modules allow for the management of institutional users, adding and deleting personnel, assigning permissions, retrieving national standards, and reviewing criteria to support medical institutions’ configuration needs. The evaluation standard module provides a summary of collected and reported data and displays the rating rules and scoring method. The interpretation of the detailed rule module allows hospital administrators to interpret national standards and display the detailed rules for each clause. The implementation plan module allows hospital administrators to formulate data collection plans based on the review terms, while the data acquisition module allows staff to perform periodic reporting operations according to assigned tasks. After data collection is finished, hospital managers can track the reporting status and calculate completion and accuracy rates.

TF-IDF algorithm-based patient symptom model

To construct a patient symptom model based on the TF-IDF algorithm, medical treatment samples from the data warehouse of the medical company can be employed. However, words used for different symptom segmentations have varied importance in describing the real condition of patients. Therefore, it is necessary to assign weights to a word describing each symptom to measure its importance in a patient’s condition. To determine the significance of a symptom participle in a patient’s chief complaint, its frequency in the patient’s complaint must be compared to its frequency in other chief complaints. A higher frequency of a symptom participle in a patient’s complaint and a lower frequency in other complaints indicates that the symptom is more important in the patient’s current symptoms.

Different types of symptom segments in the main complaint can be utilized as featured words. Then, the TF-IDF of various symptom segments of patients can be utilized as the weight of the featured words to generate the multidimensional space vector model (VSM). The construction of the patient symptom entity is thus completed by collecting those frequencies. TF-IDF algorithm quantifies the importance of a symptom participle in the patient’s chief complaint by calculating the ratio of the frequency of a symptom participle in the patient’s chief complaint to the total frequency of all symptom participles in the patient’s current chief complaint. Equation (9) presents the calculation.

(9) ${\text{TF}}_{\text{i},\text{j}}{\text{=n}}_{\text{i},\text{j}}{\text{/n}}_{\text{*},\text{j}}$where n_i,j denotes the frequency of symptom participle ti appearing in patient j’s current complaint, n_*,j represents the total number of occurrences of all symptom participle in patient j’s current complaint. IDF represents the inverse document frequency of symptom participle i in the patient j’s main complaint file. Equation (10) presents the IDF calculation.

(10) $\mathrm{I}\mathrm{D}{\mathrm{F}}_{\mathrm{i}}=\mathrm{l}\mathrm{o}\mathrm{g}(\mathrm{N}/{\mathrm{D}}_{\mathrm{i}}+1)$

The condition of patient j can be represented by the VSM v(j) composed of n symptom featured word segments in the chief complaint of patient j. Equation (11) presents it.

(11) $\mathrm{v}(\mathrm{j})=({\mathrm{W}}_{\mathrm{j}1},{\mathrm{W}}_{\mathrm{j}2},{\mathrm{W}}_{\mathrm{j}3},\dots ,{\mathrm{W}}_{\mathrm{j}\mathrm{n}})$

Score-weighted TF-IDF-based department symptom vector

To accurately measure the value of different symptom participles for department-featured descriptions, the TF-IDF algorithm introduced earlier is implemented. However, when constructing an individual department symptom vector model, some differences occur. The patient’s symptom participle is based on the patient’s initiative expression, which is closer to the actual situation. In contrast, the latter is based on the frequency of occurrence of each symptom within the patient’s medical record. Regarding departmental triage, conventional triage methods may have erroneously assigned departments in the past, leading to misdiagnosis and incorrect treatment issues. However, by considering patients’ feedback and ratings of doctors, we can infer the suitability of a department for treating certain symptoms. If a patient visits a particular department and rates the doctor with a high score, the visited department is likely more suitable for treating the patient’s symptoms. This also implies that the set of the patient’s symptoms has a higher contribution to this department’s feature description. Conversely, a lower score indicates that the patient’s symptoms contribute less to the department’s description.

Therefore, to improve the accuracy of department triage, we need to improve the conventional TF-IDF algorithm and add the weight effect of scoring. Equation (12) presents the calculation of the score-weighted word frequency TF of symptom i to a certain department. Equation (12) presents its calculation.

(12) $$\mathrm{T}{\mathrm{F}}_{\mathrm{i}}={\displaystyle \frac{{\sum }_{k\in U}{n}_{i,k}\cdot r_k}{{\sum }_{k\in U}{n}_{\ast ,k}\cdot r_k}}$$where n_i,k represents the frequency of symptom i in the chief complaint symptom set of patient k in department d, and r_k (1 ≤ r_k ≤ 5) represents the score given by the patient to the doctor after the visit is realized. N_*,k represents the total number of occurrences of all symptom words in the chief complaint symptom set of patient k. U denotes the collection of all patients in department d. The department d can be represented by the contribution of its n symptoms forming an eigenvector space model v(d) calculated by Eq. (13).

(13) $\mathrm{v}(\mathrm{d})=({\mathrm{W}}_{\mathrm{d}1},{\mathrm{W}}_{\mathrm{d}2},{\mathrm{W}}_{\mathrm{d}3},\dots ,{\mathrm{W}}_{\mathrm{d}\mathrm{n}})$

Problem analysis and process design

After running a department triage, the intelligent triage system recommends doctors to patients by utilizing collaborative filtering based on neighboring users. However, as the number of patients grows largely, finding similar patients becomes time-consuming, so it requires performing similarity calculations with each patient and finding the k most similar patients. To address this issue, historical patients belonging to a given department can be aggregated into classes by implementing an improved K-means algorithm by utilizing trust relationships. The optimal number of clusters for each department and their centroids are then determined to reduce computational burden during real-time implementations.

After clustering the collection of historical patients of a department, when a new target patient is input based on his chief complaint, the department to which he is assigned is determined, and the distance between the patients and the centroid of each cluster of the department is compared to determine which cluster the patient belongs to. This two-layer screening enables the department to cluster patients quickly and find the k neighbor patient-doctor score matrices that are most similar to the target patient, reducing the computational complexity of collaborative filtering. Department patient clustering supports the next step of the collaborative filtering of the doctor recommendation process based on matrix decomposition and user similarity.

Improved K-means algorithm based on trust relationship

Conventional recommendation algorithms face challenges with increasing amounts of user data, when the similarity between users is calculated, which becomes time-consuming and resource-intensive. To address this issue, researchers have developed improved algorithms that handle big data more efficiently. For instance, K-means clustering can be implemented to split user data into different clusters based on similarity, reducing the number of irrelevant users in the similarity calculation. However, employing Euclidean distance-based methods may lead to inaccurate patient clusters. In addition to symptom feature word weights, the degree of trust among different patients is also significant in determining the accuracy of models.

By incorporating K-means clustering into the hybrid recommender framework, the accuracy and efficiency of the department triage and doctor recommendation processes are enhanced. So, it enables the system to identify relevant patient clusters and recommend appropriate departments and doctors based on similarity and patient preferences. Specifically, K-means clustering is known for its scalability and efficiency in handling large datasets. In the context of medical appointment platforms, where there may be a considerable number of patients and doctors, the K-means clustering allows for effective and scalable patient grouping and recommendation processes. Also, it enables the system to handle a significant volume of patient data and provide timely and accurate recommendations. In addition, K-means clustering generates well-defined and interpretable clusters, making it easier to understand and interpret the patient groups. The resulting clusters can provide valuable insights into patient characteristics, allowing medical institutions to tailor their services and resources accordingly. Furthermore, it facilitates the identification of patterns and trends within patient data, leading to more informed decision-making.

The trust relationship is defined as the number of symptom-characteristic words that appear in two patients’ complaints, indicating a higher possibility of suffering from the same disease. The degree of trust is calculated by Eq. (14).

(14) $$\mathrm{T}\mathrm{u},\mathrm{v}={\displaystyle \frac{\sqrt{I\left(u\right)I\left(V\right)}}{I\left(u\right)I\left(V\right)}}.$$

The collections of symptoms of patients u and v are represented by I(U) and I(V), respectively. T_u,v represents the proportion of symptoms that co-occur in the two chief complaints to all symptoms of the two patients. A higher value indicates greater similarity between the diseases the patients suffer from, leading to higher symptom confidence. Therefore, when the K-means algorithm is employed based on the trust relationship, the distance between two patient representations should be calculated accordingly, and the centroid should be determined by Eq. (15).

(15) $${\mathrm{D}}_{\mathrm{u},\mathrm{v}}={\mathrm{T}}_{\mathrm{u},\mathrm{v}}\cdot \sqrt{{\sum }_{i\in I\left(u,v\right)}{\left(W_{u,i}-W_{v,i}\right)}^2}.$$

The weight value of the symptom feature word i in the chief complaint of patient u and patient v is represented by W_u,v and W_v,i, respectively. I(u,v) denotes the set of common symptoms between patients u and v. Figure 6 illustrates the algorithm.

Problem analysis and process design

To reduce calculation time and determine the most similar m-nearest neighbor patients to the target patient, clustering is utilized. However, since most patients do not visit all doctors in the department and score them at once, the resulting most similar m-neighbor patient-doctor score matrix is sparse. This can negatively impact the prediction of some doctors’ scores affected by only a few neighboring patients, lowering the recommendation effect. To address this, SparkMLlib’s ALS matrix decomposition is employed to transform the scoring matrix into a dense, similarly structured score matrix. Using collaborative filtering based on user similarity score, the predicted score of the target patient for the doctor is obtained, with the TopN predicted score as the final recommendation list.

Matrix decomposition of neighboring patients based on ALS

After predicting the target patient’s department based on clustering of historical patient data, the most similar m-patients with the same symptom cluster are identified along with n doctors that they have scored, constructing a patient-to-doctor score matrix R_m×n. However, this matrix is often very sparse due to many empty entries. Although employing user-based collaborative filtering directly can predict doctors’ ratings, it may result in inaccurate predictions due to a small number of similar patients that rate doctors.

SparkMLlib can transform the patient-doctor score matrix to the inner product of two low-dimensional, small matrices P and Q, respectively. Each with its corresponding hidden factor is presented. The ALS algorithm assumes that the approximate patient-doctor scoring matrix R_m×n has a low-rank scoring matrix $\tilde{\mathrm{R}}$_m×n, which can be approximated by employing two small matrices P_(k×m) and Q_(n×k), respectively.

Predictions of rating and recommendation

Assuming that the sparse scoring matrix of similar neighbor patients-doctors for the existing target patient a is R_m×n and $\tilde{\mathrm{R}}$_m×n is the approximate scoring matrix obtained after matrix decomposition by employing the ALS algorithm. When compared to R_m×n, $\tilde{\mathrm{R}}$_m×n is dense, which avoids the accuracy reduction of recommendations due to sparsity in the adjacent patient matrix. The score prediction of the target patient concerning the doctor is determined based on neighbor patient scores in $\tilde{\mathrm{R}}$_m×n and the similarity between symptoms of the target patient a and those of neighboring patients. Equation (16) presents this calculation.

(16) $${\tilde{r}}_{ai}={\overline{r}}_i+{\displaystyle \frac{{\sum }_{j\in N_{\left(a\right)}}Sima\left(a,j\right)\cdot \left({\tilde{r}}_{ji}-{\overline{r}}_j\right)}{{\sum }_{j\in N_{\left(a\right)}}Sima\left(a,j\right)}}$$where $\tilde{\mathrm{r}}$_i denotes the average score of doctor i in the approximate scoring matrix, $\tilde{\mathrm{R}}$_m×n is obtained by matrix decomposition. N_(a) represents the set comprising target patient a and m neighbor patients in the cluster. Sim_(a,j) represents the similarity of symptom vectors between target patient a and neighboring patient j, $\tilde{\mathrm{r}}$_ji denotes the score of doctor i by similar patient j of the target patient a in the approximate scoring matrix $\tilde{\mathrm{R}}$_m×n, $\tilde{\mathrm{r}}$_j denotes the average rating of patient j to the doctor in the approximate rating matrix $\tilde{\mathrm{R}}$_m×n. The score prediction of the target patient a on n doctors who have scored neighbor patients can be computed with the highest score.