Formalization and acquisition of temporal knowledge for decision support in medical processes

Highlights

• A full classification of time constraints for medical procedures is proposed.

• Two representation formalisms are proposed and validated to represent time constraints.

• Three algorithms for automatic generation of generalized time constraints from clinical databases are proposed.

• Time constraints were generated automatically from the data about Arterial Hypertension patients.

• The study on adherence of generated time constraints to CPG recommendations is provided.

Abstract

Background: In medical practice, long term interventions are common and they require timely planning of the involved processes. Unfortunately, evidence-based statements about time are hard to find in Clinical Practice Guidelines (CPGs) and in other sources of medical knowledge. At the same time, health care centers use medical records and information systems to register data about clinical processes and patients, including time information about the encounters, prescriptions, and other clinical actions. Consequently, medical records and health care information systems are promising sources of data from which we can detect temporal medical knowledge.

Objective: The objectives were to (1) Analyze and classify the sorts of time constraints in medical processes, (2) Propose a formalism to represent these sorts of clinical time constraints, (3) Use these formalisms to enable the automatic generation of temporal models from clinical data, and (4) Study the adherence of these intervention models to CPG recommendations.

Methods: In order to achieve these objectives, we carried out four studies: The identification of the sort of times involved in the long-term diagnostic and therapeutic medical procedures of fifty patients, the supervision of the indications about time contained in six CPGs on chronic diseases, the study of the time structures of two standard data models, as well as ten languages to computerize CPGs. Based on the provided studies, we synthesized two representation formalisms: Micro- and macro-temporality. We developed three algorithms for automatic generation of generalized time constraints in the form of micro- and macro-temporalities from clinical databases, which were double tested.

Results: A full classification of time constraints for medical procedures is proposed. Two formalisms called micro- and macro-temporality are introduced and validated to represent these time constraints. Time constraints were generated automatically from the data about 8781 Arterial Hypertension (AH) patients. The generated macro-temporalities restricted visits to be between 1–7 weeks, whereas CPGs recommend 2–4 weeks. Micro-temporal constraints on drug-dosage therapies distinguished between the initial dosage and the target dosage, with visits every 1–6 weeks, and 2–5 months, respectively. Our algorithms obtained semi-complete maps of dosage increments and the maximum dosages for 7 drug types. Data-based time limits for lifestyle change counsels and blood pressure (BP) check-ups were fixed to 6 and 3 months, for patients with low- and high-BP, respectively, when CPGs specify a general 3–6 month range.

Conclusions: Experience-based temporal knowledge detected using our algorithms complements the evidence-based knowledge about clinical procedures contained in the CPGs. Our temporal model is simple and highly descriptive when dealing with general or specific time constraints’ representations, offering temporal knowledge representation of varying detail. Therefore, it is capable of capturing all the temporal knowledge we can find in medical procedures, when dealing with chronic diseases. With our model and algorithms, an adherence analysis emerges naturally to detect CPG-compliant interventions, but also deviations whose causes and possible rationales can call into question CPG recommendations (e.g., our analysis of AH patients showed that the time between visits recommended by CPGs were too long for a proper drug therapy decision, dosage titration, or general follow-up).

Introduction

Time is intrinsic to medical practice. Therefore, time must be part of any representation of medical practice knowledge or only instantaneous clinical decisions could be made. Long-term interventions and planning in medicine are common and they demand a description of time constraints about the clinical actions involved. These actions are correct not only if they are done, but also if they are done at and during the right time. The health conditions of patients start and hold during a set time or over a time period. Clinical interventions occur at one or more points in time or over periods of time [42]. However, this temporal dimension is not always available, or is just partially considered, in Clinical Practice Guidelines (CPGs). CPGs are conceived to support decision-making processes in health care, covering health maintenance, prevention, diagnosis, treatment, patient self-care, and education. They describe medical procedures as a narrative set of recommendations for the management of patients who have a particular disease. CPGs dealing with chronic disorders require the patient’s treatment to be continued over time, often for the remainder of the patient’s life [6]. Some CPGs can include certain time-related parameters, such as the recommended time between two consecutive encounters during dosage titration. Nevertheless, very often, times are unspecified and, consequently, physicians are supposed to fill these information gaps with their own knowledge and experience.

Several studies have shown that the quality of health care is directly related to the experience of the physicians involved [8], [24], [25], [32]. The argued rationale is that physicians with experience can reason better than less experienced or inexperienced physicians because they have a greater ability to combine different sorts of knowledge acquired from several sources and from their own professional education, training, and experience. Consequently, they use to make wiser decisions. However, even senior physicians may have serious difficulties, or they might be reluctant to ascertain general time constraints for some diseases or clinical actions. To our knowledge, currently, there are no mechanisms to help physicians obtain evidence-based knowledge about the time constraints that should be included in CPGs. However, health care centers use medical records and information systems to register data about clinical processes and patients, including information about the times in which the encounters took place, the prescriptions that were made, or when other clinical actions happened [3], [5], [25], [27], [29], [30], [47], [48]. Information overload is becoming an issue as patients’ data and medical information are being accumulated. We are heading towards an era of big data, where the issue is not only the amount of data available, but also the velocity of accumulation and sharing, as well as its variety. Many issues appear related to this data explosion, such as the complexity and heterogeneity of clinical data. The interest in discovering hidden information is emerging [28]. This information is not necessarily evidence-based, but gathered from the experience of daily medical practice. It can be recognized and transformed into medical knowledge with the help of computer intelligent data-analysis techniques and tools to support decision-making [3], [4], [5], [27], [29], [30], [47], [48], [49]. Consequently, medical records and health care information systems are promising sources of information to detect medical knowledge about time in medicine. When modelled, this experience-based knowledge about time can complement the evidence-based knowledge about clinical procedures contained in the CPGs, thus allowing Timed Medical Decision Support Systems’ (TMDSS) development and exploitation.

In [21], Extended Timed Transition Diagrams (eTTDs) were proposed as an intuitive, easy, and efficient mechanism to formalize knowledge about medical procedures. eTTDs are state-transition diagrams in which states describe clinical situations in terms of a set of state variables with a clinical sense. Transitions between pairs of eTTD states are possible. They are represented as state-to-state edges that are labelled with (1) Clinical conditions that the patients crossing the edge satisfy, (2) Medical activities or actions describing the treatment provided to these patients, and (3) A time interval constraint representing the time for them to evolve to the next state. An eTTD can have transitions showing clinical situations in which a patient’s condition remains the same following a clinical activity and time interval. Clinical actions in the transitions that correspond to pharmacological interventions can include a dose (e.g., 200 mg) and an intake frequency (e.g., every 8 hours). Both the dose and the intake frequency define the dosage of the pharmacological action.

In [21], eTTDs were used to represent knowledge about medical practice for the management of chronic patients at three levels of detail: Therapy strategy, dosage titration, and drug intolerance management. The modelling of these three levels of clinical knowledge was applied to offer help in decision-making processes for long-term treatments. The main limitation of that work was the lack of formalized representation for time constraints.

Here, in order to complement that work and to identify common time constraints in medical practice, we analyzed diagnostic and therapeutic medical procedures for 50 randomly chosen patients registered in the primary care information system of the SAGESSA Health Care Group¹ (Spain) [41]. We also analyzed several CPGs about chronic diseases in order to detect the sorts of time constraints contained, and studied the time constructs of the standard data models EHRcom [11] and CDA HL7 [19], as well as several languages to computerize CPGs, to detect the sort of time constraints that they can represent. Based on this quadruple analysis, we synthesized two representation formalisms of time constraints in medicine: micro- and macro-temporality. Here, we also provide algorithms to generalize time constraints in the form of micro- and macro-temporalities from hospital datasets. When detected, constraints can be used to (1) complement evidence-based knowledge from CPGs with experience-based temporal knowledge obtained from data and to (2) study the adherence of physicians to the time constraints suggested in CPGs.

The rest of this paper is organized as follows. Section 2 deals with our proposal of time modelling for medical procedures. It begins with a classification of the time constraints for medical processes that we found after our analyses were conducted. It continues with a proposal on modelling the temporal dimension of medical procedures, and provides validation of the proposed temporal model. In Section 3, we propose computer algorithms for automatic generation of time constraints derived from data on medical processes. It also provides two types of tests and a discussion on the obtained results considering their medical correctness and adherence. Discussion and conclusions are presented in Section 4 and Section 5, respectively.