Method and computer system for data transfer during a confidential consultation chat

Abstract

The invention relates to a method for data transfer during a confidential consultation chat between an adviser unit (T) and at least one client unit (P). The method includes a formulation (S1.1) of a chat message at the adviser unit (T) with assistance from a first large language model (LLM1), which is executed on the adviser unit (T). Further, the method includes an encryption (S2) of the chat message formulated by the adviser unit (T) and a creation (S3.1) of a cyber physical social contract, CPSC, by the adviser unit (T). Further, the method includes a transfer (S3) of a first dataset from the adviser unit (T) to the client unit (P), the first dataset containing the encrypted chat message from the adviser unit (T). Moreover, a computer program comprising commands is described, said commands, upon execution of the computer program by a computer system (2), prompting said computer system (2) to execute such a method. Further, a computer system (2) is described, said computer system being configured to execute such a method and / or such a computer program.

Description

[0001] Method and computer system for data transmission in a confidential consultation chat

[0002] Technical area

[0003] The present invention relates to a method for data transmission in a confidential consultation chat between a supervisor unit and at least one client unit. The present invention further relates to a computer program and a computer system comprising a supervisor unit and at least one client unit.

[0004] State of the art

[0005] Methods and systems are known from the prior art to enable data transmission during a psychotherapeutic chat session. Such methods and systems can also be used to transmit chat messages between a therapist and a patient. Furthermore, it is known from the prior art to use automatic recognition to identify characteristics of the therapist and, alternatively or additionally, of the patient from these transmitted chat messages and to inform the therapist of these characteristics. Further therapy can be carried out by the therapist based on this information.

[0006] In general, however, electronic messages cannot meet the personal or legal requirements of psychotherapy, which is particularly sensitive to patient data protection. Therefore, patient acceptance may be low, or legal regulations may prevent the electronic exchange of chat messages in psychotherapy.

[0007] Psychotherapists are also highly trained specialists. The use of electronic chat messages can potentially increase the efficiency of therapy sessions. It is desirable for psychotherapists to be able to provide effective care to as many patients as possible, thus enabling cost-effective access to psychotherapy.

[0008] Description of the invention

[0009] A first aspect relates to a method for data transmission in a confidential counseling chat, such as a counseling text chat, between a supervisor unit and at least one client unit. The confidential counseling chat can be a psychotherapeutic chat session. The counseling chat can have messages in text form, in audio form, in image form and / or in video form. The supervisor unit can be a coach unit, a counselor unit or a therapist unit. The client unit can be a patient unit. For example, a chat message can consist only of text or have at least a text portion. The confidential counseling chat can be a psychotherapeutic chat session. The supervisor can be, for example, a coach or a therapist, in particular a psychotherapist. The client can be a person to be coached or a patient. The method can be a computer-implemented method.The data transfer can be a data exchange and, alternatively or additionally, communication between the supervisor unit and the client unit. The method can also be designed to generate parts of the transferred data itself, for example to relieve the supervisor. The method can be set up for data transfer between exactly one supervisor unit and exactly one client unit and also for data transfer between one or more supervisor units and one or more client units. The data transfer can take place bilaterally between a supervisor unit and a single client unit. For example, the supervisor unit can have a chat session with a first client unit and another chat session with a second client unit. However, the two chat sessions can be independent of each other.For example, the data transmission and / or data generation by the method can occur separately for each chat session. The supervisor unit can be, for example, a computer or a smartphone. The client unit can also be, for example, a computer or smartphone. A user of the supervisor unit can be a supervisor. A user of the consultant unit can be a consultant. A user of the coach unit can be a coach. A user of the therapist unit can be a therapist, in particular a psychotherapist. A user of the client unit can be a client, and a user of the patient unit can be a patient.

[0010] The method comprises formulating a chat message on the supervisor unit with the support of a first large language model, LLM, which is executed on the supervisor unit. Formulating with the support of the first large language model can be an auxiliary formulation of the chat message by the first large language model. For example, supporting the first large language model can comprise pre-formulating the chat message or a text for sending by the supervisor. A user of the supervisor unit, for example a supervisor, can use this pre-formulated text to formulate the chat message. For example, they can accept this pre-formulated text as it is and release it for sending as a chat message. Alternatively, they can partially accept this pre-formulated text to formulate the chat message.For example, the supervisor can partially or completely modify the pre-formulated text by manually entering it into the supervisor unit. The supervisor can reject the pre-formulated text and formulate a chat message entirely manually. The first large language model can be a language model based on artificial intelligence. The chat session can feature an instant messaging system. The first large language model can, for example, use previously exchanged chat messages as input, for example from the current or previous chat session. The first large language model can, for example, have been trained with chat messages from counseling chats, whereby these counseling chats can be taken from real counseling chats, can be artificially generated and / or can partially be pre-formulated or generated chat messages by another large language model.

[0011] By supporting the Large Language Model, the therapist can be relieved of the burden of generating chat messages during a therapy session. For example, a client can pre-formulate a therapeutically meaningful message, which the therapist only needs to slightly adapt to the individual patient. This can significantly reduce the therapist's workload, even making it possible to care for multiple patients simultaneously. Furthermore, the Large Language Model can also simplify therapist compliance with therapeutic standards and / or regulations with which the Large Language Model was trained.

[0012] The method further comprises encrypting the chat message formulated by the supervisor unit. The encryption can be bilateral encryption, for example, encryption such that only the supervisor unit and client unit participating in the bilateral data transmission have access to the encrypted chat message and / or can decrypt the chat message. The encryption can be asymmetric encryption, for example, using a public key. For example, a public key of the client unit may then be necessary for encryption. A private key matching this public key of the client unit may only be known to the client unit and / or a client at the client unit. This private key can be used to decrypt the encrypted chat message.

[0013] The method further comprises creating a Cyber ​​Physical Social Contract (CPSC) by the supervisor unit. In the following description, the abbreviation CPSC stands for Cyber ​​Physical Social Contract. The CPSC can be a user-based contract, for example relating to the supervisor using the supervisor unit and / or the client using the client unit. The CPSC can be a legally valid contract for chat messages from the supervisor unit, in particular for their transmission to the client unit. A CPSC can be created for the formulated chat message. The creation of a CPSC by the supervisor unit comprises, for example, creating a data protection-compliant contract between the supervisor using the supervisor unit and the client using the client unit.The CPSC can also be designed as a unilateral contract by the sender of a chat message, which becomes effective upon transmission. For example, the CPSC can consent to the transmission and receipt of chat messages. The CPSC can also restrict or permit further use of the transmitted chat message, such as forwarding it to third parties and / or storing it by the sender and / or recipient.

[0014] The method further comprises transmitting a first data set from the supervisor unit to the client unit. The transmission can comprise sending and / or receiving the first data set. A secure communication channel can be used for the transmission. For example, a secure communication channel can be used in addition to encrypting the formulated chat message before transmission. The first data set comprises the encrypted chat message from the supervisor unit. The first data set can comprise further data, such as the created CPSC. The steps comprised by the method of formulating the chat message, encrypting the formulated chat message, creating the CPSC, and transmitting the first data set can be carried out by the method in the above-mentioned order. Alternatively, a different order for carrying out the steps can also be used.

[0015] The private and public keys, which form an asymmetric key pair, for example, can be fixed and thus not changeable over time. Alternatively, the asymmetric key pair can be variable. For example, a new asymmetric key pair can be created for each consultation chat, each created CPSC, and / or each transmitted data set, i.e., for each transmitted chat message. For example, a new asymmetric key pair can be created for the supervisor unit and / or for the client unit for each transmitted chat message, and can differ from previous asymmetric key pairs. Alternatively or in addition to asymmetric encryption, symmetric encryption can be used. For this purpose, for example, both the supervisor unit and the client unit can each have a private key and a corresponding public key.A shared secret key can then be created using the Deffie-Hellman key exchange for symmetric encryption. This shared secret key can then be used to encrypt and decrypt chat messages, with the shared secret key being available on both the host unit and the client unit, and only available there. For example, the created CPSC can be closed using the Deffie-Hellman key exchange, where the key exchange is considered a declaration of intent. If the respective key pairs on the host and client units are regularly regenerated, and a new shared secret key is regularly generated based on this via key exchange, this further increases the security of the process.Such symmetric encryption increases security against third parties intercepting the chat message.

[0016] This method thus enables the implementation of confidential counseling chats, such as private psychotherapy chat sessions or coaching chat sessions, with client support. The created CPSC enables a data protection-compliant and legally valid confidential counseling chat. At the same time, by formulating the chat message with support from the first Large Language Model, pre-formulation of texts with client support is possible. This can support the caregiver using the caregiver unit in formulating chat messages. By encrypting the formulated message before transmission, a secure transmission of the first data set with the chat message from the caregiver unit to the client unit can be ensured.This method can be used to implement a language-based pre-formulation using the Large Language Model, which supports the formulation process by the Large Language Model running on the support unit and is thus local. This local execution technically simplifies and accelerates the formulation and support process. Furthermore, it can make it more difficult for unauthorized third parties to access confidential client communication, which might otherwise be easily possible with a Large Language Model running on a central server.The method can thus be used for data transmission in a bilateral confidential counseling chat, such as a psychotherapeutic chat session, wherein the counselor is supported by means of the client in formulating the chat message, a data protection-compliant contract is created, and the chat message is encrypted before transmission in order to ensure data security and a legally compliant electronic counseling chat, such as a therapy session. According to a further embodiment, the method can further comprise formulating a chat message on the client unit. A client can formulate and compose the chat message manually. The formulation of the chat message can be carried out locally on the client unit, for example by typing or voice commands. The method can further comprise encryption of the chat message formulated by the client unit.Encryption can be performed using a public key of the supervisor or the supervisor unit. A matching private key of the supervisor can then be used later to decrypt the encrypted chat message.

[0017] Furthermore, the method can comprise the creation of an additional CPSC by the client unit. This additional CPSC can also be a user-based contract, in this case based on the client and / or the supervisor. This CPSC can also be a legally valid contract for the chat messages, in particular their transmission from the client unit to the supervisor unit. Optionally, the creation of the additional CPSC can be omitted, and the chat messages created on the client unit can already be covered by the CPSC created on the supervisor unit. If a chat message is first created on the client unit, the CPSC created there can also cover chat messages from the supervisor unit, and the CPSC created there is no longer necessary.

[0018] Furthermore, the method can comprise transmitting a second data set from the client unit to the caregiver unit. The second data set can be different from the first data set. The second data set can comprise the encrypted chat message from the client unit. With such a method, a client using the client unit can manually formulate chat messages. This chat message formulated by the client can then be encrypted, for example, before its transmission. Then, for a data protection-compliant confidential counseling chat, such as a psychotherapeutic chat session, another CPSC can be created by the client unit. Then, a second data set comprising the encrypted chat message from the client unit can be transmitted from the client unit to the caregiver unit, such that the chat message can be transmitted from the client to the caregiver.The second data set may also contain additional data, such as the additional CPSC created.

[0019] According to a further embodiment, the formulation of a further chat message by the caregiver unit can be carried out with the support of the first large language model depending on the chat message transmitted from the client unit. Further data from the second data set can also be taken into account. Respective second data sets transmitted from the client unit to the caregiver unit can be an input for the large language model in supporting the formulation of further chat messages at the caregiver unit. In addition, chat messages formulated at the caregiver unit can also serve as input. For example, an entire chat history from a current and / or previous consultation chat or therapy session with the same client and optionally the same caregiver can form an input. The first large language model can be a special large language model.In general, a large language model can have an input and an output. For example, one input of the large language model can be information from a chat message. Information about a chat message as input to the large language model can, for example, be information about the second data set, such as the encrypted chat message from the client unit. An output of the large language model can be a pre-formulated text for formulating a chat message from a supervisor unit. Thus, with the method and the first large language model, a chat message from the client can be responded to and, based on this, a text can be pre-formulated with client support to assist in formulating the chat message to the supervisor unit.

[0020] According to a further embodiment, the first large language model can be stored locally on the support unit. The first large language model can be stored on a memory of the support unit using the method, for example, after the large language model has been transferred from a central server and / or after the large language model has been trained locally. This protects the large language model from manipulation and can also be available at any time to support the formulation of chat messages on the support unit.

[0021] Support through the first large language model can be performed locally on the supervisor unit. For this purpose, the method can retrieve or read the first large language model stored in a memory of the supervisor unit. Alternatively, the first large language model can also be loaded temporarily onto the supervisor unit. By performing support locally through the first large language model, the chat message can be formulated locally and on the supervisor unit, even with client support. For example, there is no need for a constant data connection to another computing device, such as a server, on which a large language model could otherwise be stored and executed to support formulation. Furthermore, there is no need to transfer sensitive data to third parties, which would, for example, first have to be contractually regulated with the client.

[0022] According to a further embodiment, the transmitted first data set can contain the CPSC created by the supervisor unit. For example, this can lead to the completion of the CPSC. By transmitting the CPSC, it can be accepted by the client and / or stored on the client side. Alternatively or additionally, the transmitted second data set can contain the CPSC created by the client unit. This can lead to the completion of the CPSC created by the client unit. By transmitting the further CPSC, it can be accepted by the supervisor and / or stored on the supervisor side. The completion of a CPSC can be performed automatically by transmitting a data set. A CPSC therefore does not have to be completed manually between the supervisor and the client. This can accelerate and simplify the process for data transmission and the exchange of chat messages.However, the transfer can still occur based on user input from the client and / or supervisor. This can constitute a declaration of intent to conclude the contract from the client or supervisor. For example, each transfer of a chat message in an instant message can be a declaration of intent to conclude a respective microcontract for each chat message. Such a contract or CPSC can be automatically concluded and saved in the background. For example, the CPSC created by the supervisor unit is transferred to the client unit and accepted by the client unit. Acceptance of the contract can occur by receiving this CPSC created by the supervisor unit. Alternatively or additionally, the client can accept the CPSC by decrypting the chat message encrypted by the supervisor unit on the client unit.The same can apply in reverse for the additional CPSC created on the client unit.

[0023] According to a further embodiment, the method can comprise a step of locally storing transmitted chat messages. The supervisor unit and, alternatively or additionally, the client unit can comprise a memory, such as a hard drive. The transmitted chat messages can be stored in this memory. For example, encrypted or unencrypted chat messages can be stored on the local memory. Alternatively or additionally, the method can comprise a step of storing created CPSCs. For example, the CPSC created by the supervisor unit can be stored on a memory of the supervisor unit. This allows chat messages to be stored and, alternatively or additionally, created CPSCs to be stored locally on the client unit or supervisor unit and not externally, for example, on a server.Overall, each transmitted data record can be stored locally. A message history can also be created and saved from the transmitted chat messages. Storage can also be encrypted. This allows, for example, the supervisor or the client to view previously transmitted chat messages locally and retrieve information about them. Furthermore, each chat history can be made available as input for the large language model. A data connection to a server, for example, is not necessary for this. This makes the process fail-safe against a possible failure of a server on which transmitted chat messages and created CPSCs would otherwise be stored. A security-relevant attack on data is also made more difficult because the data is stored decentralized on the supervisor unit or client unit and its memories.Saved and transmitted chat messages can be retrieved by the caregiver or client, for example, to rate a consultation chat. This rating can also be saved and / or used to train the large language model.

[0024] According to a further embodiment, the method may further comprise obtaining data donation permission for transmitted chat messages from the supervisor unit and the client unit. The supervisor using the supervisor unit and the client using the client unit can each grant data donation permission. The data donation permission may be a permission that allows data donation, i.e., the transmission of chat messages to third parties. The data donation permission may be granted, for example, for individual transmitted chat messages, transmitted chat messages from an entire consultation chat, such as a therapy session, or even for an entire stored chat history. The method may further comprise creating another CPSC depending on the data donation permission.The additional CPSC can be created if data donation permission has been granted by the client and supervisor, or a separate additional CPSC can be created for each data donation permission. The CPSC can be created for the data donation and, for example, contain restrictions such as a purpose for which the donated data may be used. One such purpose could be training the large language model. These additional CPSCs can be created bilaterally between the training server and the supervisor unit, as well as between the training server and the client unit. This means that one CPSC for the data donation applies between a company or person operating the training server and the supervisor or the client. A trilateral CPSC can also be created through the data donation. This CPSC for the data donation then applies between the company or person operating the training server, the supervisor, and the client.

[0025] The method may include transmitting the chat messages transmitted between the supervisor unit and the client unit, for which data donation permission has been granted, to a training server. Prior to transmission, the chat messages may be encrypted, for example, using a public key of the training server. The training server may then decrypt the encrypted chat messages using a corresponding private key. Transmission may occur, for example, only when data donation permission has been granted by both the client and the supervisor. Furthermore, the method may include training a second large language model with the training server based on the chat messages transmitted to the training server. Training may be performed using chat messages from multiple chat sessions, or alternatively, using chat messages from one chat session.If training is conducted using chat messages from different chat sessions, this can be done with different chat sessions from a pair of supervisor and client, or from different pairs of supervisor and client. The first large language model can serve as a basis, which can then be improved through training. Ethical aspects, for example, can be taken into account during training. Likewise, ratings contained in the chat messages can be transferred to the training server, and training can be carried out based on these ratings. For example, a client can rate a consultation chat, such as a chat session, and a corresponding chat message as particularly helpful, and training can then be carried out based on this rating.Likewise, any changes to chat messages pre-formulated by the first large language model can be saved by the supervisor and transferred to the training server. Training can be carried out depending on these changes. After the second large language model has been successfully trained, for example when the second large language model has been created, the transferred chat messages, i.e. the data donated by the supervisor and client, can be deleted from the training server. With this method, chat messages can be used as data donations for training a second large language model. For example, training can take place regularly, i.e. at regular intervals, or event-triggered, for example whenever chat messages have been transferred to the training server based on data donation permission from the client and supervisor.The second large language model can be understood as an update of the first large language model and then replaces it in the process. Alternatively, the chat messages can be not transmitted to a training server, and training can take place locally at the supervisor unit after data donation permission has been granted. A supervisor-specific second large language model can then be generated locally. This further reduces the likelihood of third-party access to chat messages. Furthermore, free computing power can be used locally, thus reducing network resources. During training on the central training server, the required computing power can be provided more cost-effectively, and donated data can be made available to improve the large language model for all supervisors using the process.In addition, donated data from different coaches can then easily be used for training, which means the database can be particularly large.

[0026] According to a further embodiment, the method can comprise a step of transmitting the second large language model to the supervisor unit. Before transmission, the second large language model can be encrypted, for example, using a public key of the supervisor, and this encrypted second large language model can be decrypted using a matching private key of the supervisor. For the transmission, a further CPSC can be created. This CPSC can be created bilaterally between the training server and the supervisor unit and thus apply, for example, between the company or person operating the training server and the supervisor. By receiving the second large language model with the supervisor unit, the CPSC can be completed, for example. The CPSC can be transmitted together with the second large language model.This allows the local large language model to be updated on the support unit. This can be done regularly. This ensures that an updated local copy of the large language model is always available in the support unit's memory.

[0027] According to a further embodiment, the caregiver unit can formulate a further chat message by supporting the second large language model instead of supporting a different large language model, such as the first large language model. Thus, for example, the updated large language model can be used to formulate the chat message at the caregiver unit.

[0028] According to a further embodiment, the method can comprise a step of locally training a third large language model with the supervisor unit as a function of transmitted chat messages. The first large language model or the second large language model can serve as the basis. The local training can be carried out independently of encrypting chat messages and alternatively or additionally independently of obtaining and granting data donation permission from the supervisor unit and alternatively or additionally by the client unit. Furthermore, the local training can be carried out independently of creating and completing another CPSC or can require only a single bilateral CPSC with the client. However, the local training can also already be approved, for example, by the CPSC created for transmitting chat messages.Local training, already possible with 4-bit quantization in a computer system, is thus possible locally with the caregiver unit. The caregiver unit can perform the local training, for example, depending on the transmitted chat messages stored on the caregiver unit, which concern one or more clients and also one or more consultation chats, such as therapy sessions. In this way, third-party large language models trained for each client and / or caregiver can be generated. These stored transmitted chat messages contain, for example, a written dialogue between caregiver and client, i.e., chat messages between the client unit and the caregiver unit.

[0029] According to a further embodiment, the custodian unit can formulate a further chat message by supporting the third large language model instead of supporting another large language model, such as the first or second large language model. Thus, updating and updating the large language model used by the custodian unit can be performed without encryption, without creating and completing another CPSC, and / or possibly even without obtaining and granting data donation permission. According to a further embodiment, the method can further comprise a step of transmitting the third large language model to another custodian unit.This allows the additional supervisor unit to formulate another chat message with the support of the third large language model instead of another large language model. In such an example, however, the additional supervisor unit does not have to train the third large language model locally. This allows the additional supervisor unit to transfer the third large language model to another supervisor unit. For this purpose, for example, a CPSC can be carried out between different supervisors assigned to the supervisor units. This CPSC can be transferred with the third large language model. As an alternative to or in addition to being transferred to another supervisor unit, the third large language model can be transferred to a server, such as the training server, and as an alternative or in addition to being transferred to a transfer server.The transfer server can be used to transfer information, such as data records, between the caregiver and client units.

[0030] According to a further embodiment, the method can further comprise deriving respective CPSCs from a template. The creation of the CPSC based on the template can be carried out depending on the respective data to be transmitted and the entities involved, for example users of the respective units. The derivation can, for example, be carried out automatically. For example, there can be a template for a message, a template for a data donation and a template for transmitting a large language model. The template can then be supplemented with, for example, a current date, and the persons and / or units involved and the transmitted data are specified, for example by numbering, date and time and data size. However, the CPSC can also contain the actual data to be transmitted in full.The CPSC can be encrypted for transmission, particularly if it contains the data to be transmitted. The derivation of the CPSC can depend on the specific template. For example, a CPSC specifically for a message can be derived from the template for the message, and so on. This allows a suitable legally valid contract to be derived for each information transfer between units. A template can be stored on each supervisor unit and client unit, so that the derivation for creation can be carried out locally on each client and supervisor unit. The same can apply to the training server as an additional unit. This means that the method can derive an application-based contract for bilateral data transmission between client unit and supervisor unit from a template, thereby creating the contract and thus concluding a contract.Manual and / or time-consuming CPSC creation is therefore not necessary.

[0031] According to a further embodiment, the method may comprise a step of determining a summary of a consultation chat. Determining the summary may be performed using a fourth large language model. The fourth large language model may be stored and executed locally on the supervisor unit and / or on the client unit. Determining the summary may be performed at the end of a consultation chat, for example, when the fourth large language model detects that farewell chat messages have been transmitted, or when the supervisor and / or client manually define the end of the consultation chat. Determining the summary may be performed locally in the supervisor unit and / or in the client unit. The method may further comprise a step of transmitting the determined summary. The method may comprise encrypting the determined summary before transmission.The method may include creating another CPSC for the summary. The method may include storing the determined summary, for example, in a memory of the caregiver unit and / or the client unit.

[0032] According to a further embodiment, the method comprises a step of training a Symbolic AI Model (SKIM). SKIM is used below as an abbreviation for Symbolic AI Model. SKIM can be based on symbolic artificial intelligence (symbolic AI). Symbolic AI can represent and manipulate knowledge depending on symbols and rules. The principle of SKIM is based on logical and linguistic principles and involves the development of knowledge-based systems that use logical reasoning to perform deductions from a set of premises. Symbolic AI systems are characterized by their ability to represent objects and concepts and to draw conclusions about them. They are frequently used in expert systems, natural language processing, and intelligent learning environments.A key challenge of this approach lies in the representation of knowledge characterized by ambiguity or context dependency. In this context, a Large Language Model can be used to transfer knowledge into a SKIM. The SKIM is interpretable by human users and based on comprehensible rules. This allows the knowledge contained in the model to be linked to guidelines, enabling evidence-based decision support. This and further explanations of SKIM can be found, for example, in the blog post "Symbolic Language vs. Subsymbolic Language" by JO Schneppat from June 1, 2023, available at https: / / gpt5.blog / .

[0033] For example, the SKIM can be trained depending on the trained large language model, such as the second or third LLM. During training, for example, information from the trained LLM is transferred to the SKIM. A user, for example a supervisor or a third party, can evaluate the trained SKIM and, depending on this, modify and / or correct the SKIM if necessary. Depending on this, the SKIM can be updated. For example, the training server can perform the training and updating, and the third party can perform the evaluation and correction via a user interface on the training server. Analogous to the trained second LLM, the SKIM can be transferred from the training server to a supervisor unit. For example, in one embodiment, the trained second LLM is transferred parallel to the SKIM.Furthermore, the transmitted SKIM can be saved locally in the supervisor unit. When formulating the chat message on the supervisor unit, support can first be provided by the second LLM, for example, a text can be pre-formulated by the second LLM. The text pre-formulated by the second LLM can then be compared with the knowledge of the SKIM. Depending on this, the pre-formulated text can then be adapted by the second LLM, for example. The supervisor can then accept this adapted pre-formulated text, for example, and thereby carry out and, for example, complete the formulation of the chat message. This can improve the support provided by the LLM for the supervisor, and pre-formulated texts are better suited to the consultation chat.

[0034] According to a further embodiment, the method may comprise a step of recording audio information on at least one of the caregiver unit and the client unit. At least one of the caregiver unit and the client unit may be communicatively connected to an audio recording device or may comprise such an audio recording device. The audio recording device may be configured to record audio information, for example, from the client and / or the caregiver. The audio recording device may be a device, for example, a box, with a microphone and a user interface. The user interface may comprise input and / or output options. The audio recording device may be arranged spatially close to the caregiver and / or client. For example, the audio recording device may be located on a table in front of the client and / or caregiver.The recorded audio information can include all audio information exchanged between the client and the caregiver. For example, the recorded audio information includes information about a dialogue between the caregiver and the client. The audio information can include a recorded conversation between the caregiver and the client. All audio information can be recorded on the caregiver unit and / or the client unit. The audio information can also be recorded during a remote consultation, for example, by telephone.

[0035] Furthermore, the method can comprise locally training a fifth large language model depending on the recorded audio information. For this purpose, the method can comprise a step of creating training data based on the recorded audio information. The first large language model, the second large language model, or even the third large language model can serve as the basis. The local training can be carried out independently of encryption of chat messages and alternatively or additionally independently of obtaining and granting data donation permission from the supervisor unit and alternatively or additionally by the client unit. However, the audio information can also be stored in encrypted form and / or stored and / or used for training only if data donation permission is obtained and granted.Furthermore, local training can be performed independently of the creation and completion of another CPSC or can require only a single bilateral CPSC with the client. However, local training using audio information can also be authorized, for example, by the CPSC created for transmitting chat messages. Local training using audio information can also be performed, for example, dependent on the creation and completion of another CPSC. Local training, for example, already possible with 4-bit quantization in a computer device, is possible locally with the supervisor unit and / or with the client unit.The caregiver unit and / or the client unit can perform local training, for example, as local distributed (federated learning) training, depending, for example, on the audio information recorded on the caregiver unit and / or client unit, which relates to one or more clients and also one or more consultation chats, such as therapy sessions. This allows for the generation of fifth large language models trained individually for each client and / or caregiver.

[0036] According to a further embodiment, the host unit can formulate another chat message by supporting the fifth large language model instead of supporting another large language model, such as the first, second, or third large language model. Thus, updating and updating the large language model used by the host unit and / or client unit can be performed without encryption, optionally even without creating and completing another CPSC, and / or possibly even without obtaining and granting data donation permission.

[0037] According to a further embodiment, the method may further comprise a step of transmitting the fifth large language model to a further supervisor unit. This allows the further supervisor unit to formulate a further chat message by assisting the fifth large language model instead of assisting another large language model. However, in such an example, the further supervisor unit does not have to train the fifth large language model locally. Thus, federated training of a large language model can be realized by one supervisor unit performing the training and another supervisor unit using the trained large language model. Thus, a model donation of the fifth large language model can be performed from the supervisor unit to the further supervisor unit.For example, a CPSC can be used between different supervisors assigned to the supervisor units. This CPSC can be transmitted using the fifth large language model. The fifth large language model can be transmitted to a server, such as the training server, as an alternative or in addition to being transmitted to another supervisor unit, and alternatively or additionally to a transmission server. The transmission server can be used to transmit information, such as data records, between the supervisor and client units.

[0038] The step of determining a summary using the fourth large language model can be performed based on the recorded audio information. Alternatively or additionally, the step of determining the summary using the fourth large language model can be performed based on chat messages.

[0039] Furthermore, training a large language model can be performed as federated learning on multiple supervisor units. This means that one large language model can be trained on one supervisor unit, and another large language model can be trained on another supervisor unit. Furthermore, a first part of training a large language model can be performed on a first supervisor unit, and a second part of training the same large language model on a second supervisor unit other than the first. The partially trained large language model can be sent from the first to the second supervisor unit. This eliminates the need to send raw data in the form of chat messages, for example, between supervisor units. This can improve data protection.A second aspect relates to a computer program comprising instructions which, when executed by a computer system, cause the computer system to execute a method according to an embodiment of the first aspect. Respective further features, embodiments, and advantages can be found in the descriptions of the first aspect. Conversely, features, embodiments, and advantages of the second aspect also represent features, embodiments, and advantages of the first aspect. The computer system can be a computer device, such as a PC, or a system consisting of several individual PCs. The computer system can, for example, be a network of distributed computers comprising servers, such as the transmission server and the training server. The computer system can also comprise the client unit and the supervisor unit. The computer system can be designed according to the third aspect described below.

[0040] A third aspect relates to a computer system with a supervisor unit and at least one client unit. The computer system can also have a training server as a further unit. The units can be connected to one another, for example, via a data connection provided, for example, by the Internet and / or a transmission server or a direct data line. The computer system is configured to execute steps of the method according to an embodiment of the first aspect and, alternatively or additionally, the computer program according to an embodiment of the second aspect. Respective further features, embodiments, and advantages can be found in the descriptions of the first and second aspects. Conversely, features, embodiments, and advantages of the third aspect also represent features, embodiments, and advantages of the first and second aspects, respectively.

[0041] According to one embodiment, the computer system can have at least two client units, wherein the data transmission can then be carried out in each case using the method according to the first aspect. The computer system can further have a transmission server and additionally a training server. The transmission server can be a single server or, alternatively, a ring structure of servers. If the transmission server is designed as a ring structure and thus as a ring network, this can lead to greater reliability.

[0042] Short description of the characters

[0043] Figure 1 shows schematically steps of a method for data transmission in a confidential counseling chat, such as a psychotherapeutic chat session.

[0044] Figure 1 a schematically shows steps of a method for data transmission in a confidential counseling chat, such as a psychotherapeutic chat session, according to an alternative embodiment.

[0045] Figure 2 shows schematically a computer system for carrying out the steps shown schematically in Figure 1, wherein a transmission server is designed as a central server.

[0046] Figure 3 shows a schematic representation of a computer system from Figure 2, in which a distributed transmission server in the form of a ring structure is used instead of a central transmission server.

[0047] Figure 4 shows a schematic of a computer system with several client units and supervisor units.

[0048] Figure 4a schematically shows a computer system with multiple client units and supervisor units according to an alternative embodiment.

[0049] Detailed description of embodiments

[0050] Figure 1 schematically shows a method for data transmission in a confidential consultation chat, such as a psychotherapeutic chat session, between a supervisor unit T and at least one client unit P. In the embodiment shown, the supervisor unit T is a therapist unit T, and the client unit P is a patient unit P. Figure 2 schematically shows a computer system 2 with a supervisor unit T and a client unit P. The computer system 2 is configured to carry out steps of the method schematically shown in Figure 1. The computer system 2 also has a training server TS and a transmission server ÜS as the central transmission server.

[0051] Furthermore, the computer system 2 has data interfaces 4 between the supervisor unit T, the client unit P, and the training server TS, each connected to the transmission server ÜS. The data interfaces 4 each form a secure communication channel. The client unit P has a memory, here embodied as a database (DBP). The supervisor unit T also has a memory (DBT). Furthermore, the supervisor unit T has a first large language model (LLM1), which is executed on the supervisor unit T.

[0052] The method comprises formulating S1.1 a chat message at the caregiver unit T with the support of a first large language model LLM1, which is executed on the caregiver unit T. The first large language model LLM1 supports a caregiver using the caregiver unit T, in this case a therapist, in formulating a chat message. This is done using artificial intelligence, Kl. The first large language model preformulates a text, and the caregiver uses this preformulated text to formulate S1.1 the chat message, for example, by accepting or modifying the preformulated text.

[0053] Furthermore, the method comprises encrypting S2 the chat message formulated at the support unit T. This is done using a public key of an asymmetric key pair of the client unit P.

[0054] The method further comprises creating S3.1 a Cyber ​​Physical Social Contract (CPSC) by the supervisor unit T. In doing so, the CPSC is derived S3.0 from a template. A template for CPSCs, or contracts, for chat messages is available in the DBT memory of the supervisor unit T. A CPSC for chat messages can be derived from this. The template can be a template from which the CPSC is created by additions during the derivation process. In the example shown, at least the supervisor and the client to whom the chat message is to be sent are taken into account when deriving the actual CPSC from the template.

[0055] The method further comprises transmitting S3 a first data set from the caregiver unit T to the client unit P. The first data set is transmitted via interface 4 and the transmission server ÜS. The first data set comprises the encrypted chat message from the caregiver unit T and, optionally, the CPSC created for it. The client, in this case a patient, using the client unit P thus receives the chat message and, optionally, the associated CPSC. Using the client's private key, the encrypted chat message can then be decrypted with the client unit P.

[0056] The method further comprises formulating S1.2 a chat message on the client unit P. The client using the client unit P manually formulates the chat message by typing it into a user interface. The chat message formulated by the client unit P is then encrypted S2. This is done using a public key of an asymmetric key pair of the supervisor unit T. Furthermore, a further CPSC is created S3.1 by the client unit P. This is also done by deriving S3.0 from a template stored in the memory DBP of the client unit P. Furthermore, a second data set is transmitted S3 from the client unit P to the supervisor unit T. The second data set comprises the encrypted chat message from the client unit P and optionally the associated CPSC. The second data set is transmitted S3 via the interface 4 and the transmission server ÜS.

[0057] For example, first a chat message is formulated S1.1 by the supervisor at the supervisor unit T, then encrypted S2 and transmitted S3 to the client at the client unit P. Then a chat message is formulated S1.2 at the client unit P. The client using the client unit P replies in the form of a dialog, for example, to the chat message just transmitted from the supervisor unit T. The dialog is continued by transmitting this chat message formulated by the client of a client unit P to the supervisor at the supervisor unit T. Chronologically following this, the formulation S1.1 of a further chat message can be carried out by the supervisor unit T with the support of the first Large Language Model LLM1 depending on the transmitted second data set and thus on the chat message from the client unit P.Optionally, the first Large Language Model LLM1 also considers chat messages previously transmitted to the client unit. The dialogue can thus continue, and the client, trained by the first Large Language Model LLM1, can respond to the chat message from the client unit P to support the caregiver in therapy.

[0058] The first large language model LLM1 is stored locally on the supervisor unit T in the DBT memory. Support provided by the first large language model LLM1 is performed locally on the supervisor unit T. This eliminates the need for a continuous connection to the transmission server ÜS or another device to support the formulation of the chat message S1.1 on the supervisor unit T. Furthermore, the confidential chat messages initially remain local.

[0059] As previously mentioned, the transmitted first data set optionally includes the CPSC created by the supervisor unit T. For example, the chat message is formulated and encrypted at the supervisor unit T. Then, by deriving it from a template, a CPSC is created at the supervisor unit T and by the supervisor unit T. By transmitting the CPSC from the supervisor unit T to the client unit P and, for example, decrypting the also transmitted encrypted chat message at the client unit P, a declaration of intent to accept the CPSC is made by the client using the client unit P. The transmitted second data set can also optionally include the CPSC created by the client unit P, and similarly, this can be accepted by the supervisor using the supervisor unit T by submitting a declaration of intent, for example by decrypting the chat message from the client unit P.

[0060] Furthermore, transmitted chat messages can be stored locally S4. Thus, the transmitted chat messages transmitted from the supervisor unit T to the client unit P can be stored in the DBT memory. Similarly, chat messages transmitted from the client unit P to the supervisor unit T can be stored in the DBP memory. Alternatively or additionally, the created CPSCs can be stored. Thus, the CPSCs created by the client unit P can be stored in the DBP memory, and the CPSCs created by the supervisor unit T can be stored in the DBT memory. A user, such as a supervisor of the supervisor unit T, can thus retrieve the chat messages and associated CPSCs stored locally in the DBT memory at a later time.A client as a user of the client unit P can also retrieve the transmitted chat messages and associated CPSC stored in the memory DBP and evaluate them, for example, in order to evaluate a past psychotherapeutic chat session.

[0061] Furthermore, the method comprises obtaining S5 a data donation permission for transmitted chat messages from the supervisor unit T and the client unit P. Thus, a supervisor at the supervisor unit T and a client of a client unit P are asked whether transmitted chat messages may be used for a data donation. Furthermore, a further CPSC is created S6 depending on the data donation permission. For example, a CPSC is created at the supervisor unit T and a further CPSC is created at the client unit P. In one embodiment, an encryption S7 of the chat messages takes place with a public key of the training server TS. Furthermore, a transmission S8 of the chat messages transmitted between the supervisor unit T and the client unit P, for which the data donation permission was granted, takes place to the training server TS.Thus, the transmission S8 always occurs precisely when both the client at the client unit P and the supervisor at the supervisor unit T have granted data donation permission for certain transmitted chat messages. A second large language model LLM2 is then trained S9 with the training server TS depending on the chat messages transmitted to the training server TS. For this purpose, the encrypted and transmitted chat messages can be decrypted using a private key that matches the public key of the training server TS. By decrypting on the training server TS, for example, the CPSCs of the supervisor unit T and the client unit P regarding the data donation can be concluded. Furthermore, in one embodiment, the method can comprise creating S10 a further CPSC depending on the trained second large language model LLM2.In this case, a CPSC can be concluded between the training server TS and a supervisor unit T. The method can further comprise a step of transmitting S11 the second large language model LLM2 to the supervisor unit T. In this case, the second large language model LLM2 is transmitted via data interfaces 4 from the training server TS via the transmission server ÜS to the supervisor unit T. Upon receipt by the supervisor unit T, the respective CPSC is then concluded.

[0062] Formulation S1.1 of another chat message by the caregiver unit T is then performed using the second large language model LLM2 instead of the first large language model LLM1. This updates the large language model used by the caregiver unit T. For example, the updated second large language model LLM2 can take previous chat messages between the caregiver and the client into account to better pre-formulate further texts for subsequent chat messages.

[0063] Furthermore, the method can comprise a step of locally training S12 a third large language model LLM3 with the support unit T as a function of transmitted chat messages. For this purpose, transmitted chat messages are used for training S12. During local training S12, obtaining permission to donate data can optionally be omitted. Furthermore, creating an additional CPSC and encrypting chat messages can also be omitted, since no further transmission of data in the form of transmitted chat messages is necessary for local training S12. Formulating S1.1 of additional chat messages by the support unit T is then carried out with the support of a third large language model LLM3 instead of with the support of another large language model LLM1, LLM2. This results in a local model update of the currently used large language model.

[0064] Figure 3 shows an alternative embodiment of the computer system 2. Instead of the central transmission server ÜS shown in Figure 2, a ring structure consisting of several servers is shown forming the transmission server ÜS. Thus, the probability of failure of the transmission server ÜS can be reduced because a decentralized configuration of the transmission server ÜS is used.

[0065] Figure 4 shows an embodiment of a computer system 2 with a plurality of client units, here a first client unit P1, a second client unit P2 and a third client unit P3. These each have a memory DB1, DB2 and DB3. Furthermore, the computer system 2 has the training server TS with the second large language model LLM2. The second large language model LLM2 is here the global and current, as it was the most recently trained large language model. This is sent as a local copy to all training units. Furthermore, the computer system 2 of the embodiment shown in Figure 4 has two supervisor units, a first supervisor unit T1 and a second supervisor unit T2. The first supervisor unit T1 has a first large language model LLM1. The second supervisor unit T2 has a third large language model LLM3. Furthermore, the two supervisor units T1, T2 each have a memory DB4, DB5.In the embodiment shown, the locally stored large language model of the supervisor unit T1 has not yet been updated. Accordingly, the supervisor unit T1 is still using the first large language model LLM1 to support the formulation S1.1, and it has not yet been updated with the globally trained large language model LLM2 provided by the training server TS. In contrast, the second supervisor unit T2 has the third large language model LLM3. This was generated by local training S12. However, it is only available at the second supervisor unit T2. The third large language model LLM3, which was generated by local training S12, can be transmitted to the first supervisor unit T1 in an embodiment not shown for updating the local copy of the large language model.

[0066] In the embodiment shown in Figure 4, the supervisor using the first supervisor unit T1 has a first chat window for chat messages with the first client unit P1 and a further, separate chat window for chat messages with the second client unit P2. In an embodiment not shown in detail, the first supervisor unit T1 has specially trained large language models for the respective clients of the client units P1, P2. In such an embodiment, the large language model has been trained client-specifically using client-specific transmitted chat messages. The support provided by a large language model thus occurs differently when formulating S1.1 a chat message depending on the client unit P1, P2. The second supervisor unit T2 has a chat session with the third client unit P3.

[0067] The computer system 2 shown enables chat messages to be transmitted S3 between the client unit P and the supervisor unit T, chat messages to be transmitted S8 to the training server TS, and trained large language models to be transmitted S11 to a supervisor unit T. The method and the computer system 2 shown enable a decentralized implementation for data transmission during the psychotherapeutic chat session. Information such as formulated and transmitted chat messages, for example, is only available locally at the client units P, supervisor units T, and exceptionally at the training server TS for training S9. The transmission server ÜS does not contain any unencrypted, personal data, such as certain chat messages. The method and computer system 2 thus enable a decentralized implementation, which enables particularly high data security.Only anonymized data on created and completed CPSCs and transmitted information, such as chat messages, are stored on the ÜS transmission server. Thus, only an anonymized logging of the data exchange is performed by the ÜS transmission server. The ÜS transmission server therefore only contains freely accessible anonymized data on a status change regarding a chat message, a data donation, or a model update, which is logged by the ÜS transmission server.

[0068] Figures 1a and 4a each show an alternative embodiment. Only differences from the embodiments shown in Figures 1 and 4 will be discussed below.

[0069] Recording S13 of audio information can take place at at least one of the supervisor unit T and the client unit P. Audio information can thus be recorded using a corresponding audio recording device on the third client unit P3. The third client unit P3 has a microphone as the audio recording device. All audio information is recorded at the third client unit P3 using the microphone. This includes the spoken word of the respective client on the third client unit P3 and of the supervisor on the second supervisor unit T2, which, as described above, is assigned to the third client unit P3.

[0070] Furthermore, the method can comprise a step of locally training S12 a fifth large language model LLM5 depending on the recorded audio information. For this purpose, all audio information exchanged between the client at the third client unit P3 and the supervisor at the second supervisor unit T2 is used for local training S12. During local training S12, obtaining data donation permission can optionally be omitted. Furthermore, creating an additional CPSC and encrypting chat messages can also be omitted, since no further transmission of data in the form of transmitted chat messages is necessary for local training S12. Formulating S1.1 further chat messages by the third client unit P3 is then carried out by supporting the fifth large language model LLM5 instead of by supporting another large language model LLM1, LLM2, LLM3.This results in a local model update of the currently used large language model.

[0071] The fifth large language model LLM5, which has been created by local training S12 as a function of the recorded audio information, can be transmitted, in an embodiment not shown, to the second supervisor unit T2 for updating the local copy of the large language model.

[0072] Reference symbol

[0073] 2 computer system

[0074] 4 Data interface

[0075] DBP, DBT, DB1-DB5 memory

[0076] LLM1, LLM2, LLM3, LLM5 Large Language Model

[0077] P Client unit

[0078] T Caregiver unit

[0079] TS training server

[0080] ÜS transmission server

[0081] S1 .1 , S1 .2 Formulating a chat message

[0082] 52 Encrypting the formulated chat message

[0083] 53 Transferring a data record

[0084] S3.0 Deriving a CPSC from a template

[0085] S3.1 Creating the CPSC

[0086] 54 Saving transferred chat messages

[0087] 55 Obtaining data donation permission

[0088] 56 Creating another CPSC depending on the data donation permission

[0089] 57 Encrypting the transmitted chat messages for a data donation

[0090] 58 Transferring chat messages to the training server

[0091] 59 Training a second large language model with the training server

[0092] 510 Creating another CPSC depending on the trained Large Language Model

[0093] 511 Transferring the trained Large Language Model

[0094] 512 Local Training of a Large Language Model

[0095] 513 Recording audio information

Claims

Patent claims 1. A method for data transmission in a confidential consultation chat between a supervisor unit (T) and at least one client unit (P), comprising the steps: - Formulating (S1.1) a chat message on the supervisor unit (T) by supporting a first large language model (LLM1) which is executed on the supervisor unit (T); - Encrypting (S2) the chat message formulated by the caregiver unit (T); - Creation (S3.1 ) of a Cyber ​​Physical Social Contract, CPSC, by the supervisor unit (T); and - transmitting (S3) a first data set from the supervisor unit (T) to the client unit (P), wherein the first data set comprises the encrypted chat message from the supervisor unit (T).

2. The method of claim 1, further comprising the steps of: - Formulating (S1 .2) a chat message on the client unit (P); - Encrypting (S2) the chat message formulated by the client unit (P); - Creation (S3.1 ) of another CPSC by the client unit (T); and - transmitting (S3) a second data set from the client unit (P) to the supervisor unit (T), wherein the second data set comprises the encrypted chat message from the client unit (P).

3. The method according to claim 2, wherein the formulation (S1.1) of a further chat message by the supervisor unit (T) is carried out by supporting the first large language model (LLM1) in dependence on the transmitted chat message from the client unit (P).

4. Method according to one of the preceding claims, wherein the first large language model (LLM1) is stored locally on the support unit (T) and wherein the support by the first large language model (LLM1) is carried out locally on the support unit (T).

5. Method according to one of the preceding claims, wherein the transmitted first data set comprises the CPSC created by the supervisor unit (T) and / or wherein the transmitted second data set comprises the CPSC created by the client unit (P).

6. Method according to one of the preceding claims, wherein the method comprises a step of locally storing (S4) transmitted chat messages and / or the method comprises a step of storing created CPSCs.

7. The method according to any one of the preceding claims, further comprising the steps of: - Obtaining (S5) a data donation permission for transmitted chat messages from the supervisor unit (T) and the client unit (P); - Creation (S6) of another CPSC depending on the data donation permission; - transmitting (S8) the chat messages transmitted between the supervisor unit (T) and the client unit (P), for which the data donation permission has been granted, to a training server (TS); and - Training (S9) a second Large Language Model (LLM2) with the training server (TS) depending on the chat messages transmitted to the training server (TS).

8. The method according to claim 7, wherein the method comprises a step of transmitting (S11) the second large language model (LLM2) to the supervisor unit (T).

9. The method according to claim 8, wherein a formulation (S1.1) of a further chat message by the caregiver unit (T) is performed by supporting the second large language model (LLM2) instead of by supporting another large language model (LLM1).

10. The method according to any one of the preceding claims, wherein the method comprises a step of locally training (S12) a third large language model (LLM3) with the supervisor unit (T) as a function of transmitted chat messages.

11. The method according to claim 10, wherein a formulation (S1.1) of a further chat message by the caregiver unit (T) is performed by supporting the third large language model (LLM3) instead of by supporting another large language model (LLM1; LLM2).

12. The method according to any one of the preceding claims, further comprising deriving (S3.0) respective CPSCs from a template.

13. The method according to any one of the preceding claims, wherein the method comprises a step of recording (S13) audio information at at least one of the supervisor unit (T) and the client unit (P) and locally training (S12) a fifth large language model (LLM5) in dependence on the recorded audio information.

14. A computer program comprising instructions which, when the computer program is executed by a computer system (2), cause the computer system (2) to carry out the method according to one of the preceding claims.

15. Computer system (2) with a supervisor unit (T) and at least one client unit (P), wherein the computer system (2) is configured to execute steps of the method according to one of claims 1 to 13 and / or the computer program according to claim 14.

Images