1 Introduction

Human development is constant, dynamic and age-related. Child development is performed by a complex interaction between prenatal environmental, hereditary factors and postnatal biological, cultural, social and educational factors. Periodic assessments of child development indicators are important since birth, with most of them having preventive purposes or for the early diagnosis of disorders that may affect child development.

There are standardized indicators on motor, cognitive, language, functioning in daily activities, behavior and communication skills called developmental milestones that characterize the neuropsychomotor and psychosocial development of children in the first years of life [6].

This early childhood development is included on global political agenda as one of the targets in the Sustainable Development Goals (SDGs) to guarantee quality access for all. In Brazil, following the global trend, interest in promoting development in early childhood has also grown [8].

The reason for this is that preventing the causes and intervening early makes it possible to modify the effects and improve the general evolutionary results of children, promoting better quality of life and autonomy in daily tasks. Both health and education professionals, as well as parents and caregivers of children, must be careful to identify risk factors for child development in advance, and make it possible to treat problems as soon as they are detected [4].

In this context, the present work aims to develop a hybrid artificial intelligence model using first-order logic and fuzzy logic that can track signs of delay in child development in a more automated way.

2 Theoretical Reference

2.1 Development Milestones Assessment Instrument

Several instruments have already been created for monitoring the child’s development and detecting possible deviations. In the 1920s and 1930s, scales were created and adapted to meet new professional demands. The Pan American Health Organization published the Manual for Child Development Surveillance, based on the Integrated Management of Childhood Illnesses (IMCI) methodology, designed to train primary care professionals. Its objective was to systematize care, facilitating guidance for parents on their children’s development and early detection of warning signs in development, so that they could be referred for appropriate assessment and treatment.

In 1984, the Ministry of Health (MS) of Brazil created the Comprehensive Child Health Care Program (PAISC). As of 2009, the Child Health Handbook began to include the child development component based on the Manual for Surveillance of Child Development, in the IMCI Context with some adaptations [3].

The Child Health Booklet is delivered to the parents of all children born in Brazil, while still in the maternity ward. It is currently the main instrument for monitoring the child’s health, to be filled out appropriately by the health professionals who accompany the child and carried by parents/guardians in any service. In addition to providing great knowledge of health history, growth and development, the instrument favors the monitoring of healthy evolution and the acquisition of new skills, promoting early interventions and preventing future damage [5] as shown in Table 1.

Documents from the Brazilian Ministry of Health, namely the Child Health Primary Care Handbook: growth and development (2012), the Child Health Handbook (2014), the Early Stimulation Guidelines: children aged 0 to 3 years with neuropsychomotor delay (2016) and the National Policy for Comprehensive Child Health Care (2018), present several milestones in child development, as well as possible warning signs about the development that must be achieved by the child in each age group. Furthermore, these documents suggest some stimulation that can performed by those responsible as soon as any change in development is identified. It also contains guidelines for referring this child to specialized and multidisciplinary treatment [7].

Table 1. Developmental Milestones at 6 months. The table is filled in according to the markers: P = Present Milestone; A = Milestone Missing, NV = Milestone not checked.
Full size table

2.2 Artificial Intelligence Approaches

Artificial Intelligence is an area of computer science that aims to make machines perform intelligent tasks, such as learning and solving problems, in a way similar to the natural intelligence of humans and animals [9]. The growing role of intelligent systems in human society raises pertinent questions about the need to explain the behavior or outcome of these systems, in order to be able to motivate their decision and make the underlying decision process understandable to humans. At this point, subsymbolic artificial intelligence techniques, such as machine and deep learning, despite their efficiency, leave something to be desired. Symbolic techniques, based on the formal representation of knowledge and its elaboration through explicit reasoning rules, end up becoming more attractive. This is especially relevant when artificial intelligence is exploited in the context of human organizations intended to provide public services, such as healthcare, diagnostic or counseling systems [1].

Among the various approaches already developed, one of the oldest and most consolidated are Expert Systems, which are based on established rules, generating reasonable advice and suggestions for solving a problem. The use of expert system as a decision support tool is decisive in solving problems that cannot be solved based on analytical calculations [2].

However, expert systems have some disadvantages, such as the lack of ability and flexibility to adapt to changes in the environment, lack of ability to generate a creative response when there is no response, lack of ability to summarize their knowledge through an analogy and the impossibility of learning from experience, not to mention that many expert systems cannot automatically change their knowledge base, nor adjust existing rules or add new ones [11]. One of the ways to mitigate these deficiencies is a combination of expert systems with other artificial intelligence techniques. For example, fuzzy logic can be used to manage uncertainty in expert systems and solve problems that cannot be solved effectively by conventional methods [10]. The main goal of fuzzy expert systems is to use human knowledge to process uncertain and ambiguous data.

3 Materials and Methods

This research is part of a broad study, covering the child’s first seven years. However, for elucidation purposes, a section was carried out, showing only the first semester of the infant’s life. To preserve children’s data and for methodological presentation purposes, data from fictitious children were artificially generated.

The proposed model is divided into two main modules, one based on first-order logic, built with the Prolog language, using the Pytholog tool, and the other based on fuzzy logic, built with the Python language, using the SciKit-Fuzzy framework.

The process of classifying the child and their development is conducted by a health professional, father, mother or caregiver. In this study, the first step, which is currently manual, can also be performed by the proposed software, with the child’s correct data being provided in advance. The process occurs according to the following steps:

  • Find the child’s age range in the age columns (in months) of the Developmental Milestones table (similar to Table 1);

  • Find the four colored lines of the same color corresponding to the Developmental Milestones of the age group;

  • Check the presence of Development Milestones or skills;

  • Fill in the corresponding spaces using the following commands: P = Present milestone; A = Missing milestone, NV = Unverified milestone.

This process is performed until the end of the child’s age group. The child is expected to have reached all anticipated milestones. But, if they have not yet reached a milestone in their age group, then an alternative process is performed, following the steps:

  • Go to previous age group;

  • Check whether the child meets the milestones in the previous range;

  • Consult the Child Classification Instrument (Table 2);

  • Classify the child’s development and adopt the indicated behavior.

Table 2. Child Development Assessment Instrument
Full size table

In this work, only Development Milestones were considered, disregarding phenotypic or physical verifications. This is because the presence of a health professional would be necessary to assess it. Therefore, this software does not produce a diagnosis, but only a partial assessment of the child, according to the defined Development Milestones.

All codes can be found at: < https://github.com/Projeto-SDIA >.

3.1 Proposed Artificial Intelligence Approach

In this article, a software is proposed as an alternative to the traditional booklet, offering a more efficient and practical approach to monitoring child development. It was based on the 2022 Children’s Handbook, as a guide for implementing the Artificial Intelligence model, focusing specifically on Development Milestones.

First Order Logic Modeling. During the process of classifying the child’s record book, the child’s development is checked based on well-defined rules. This fits into inference models consistent with first-order logic that use rules and quantifiers. To enable classification comparable to that of the booklet, using the categories ADEQUATE (AD), ALERT (AL) and DELAYED (AT), we continued by performing an empirical generation of fictitious examples of children categorized into three distinct groups: delayed development, development in alert and appropriate development. These examples, as shown below, will serve as a reference to demonstrate that the first part of the artificial intelligence model can help identify possible developmental delays, allowing for early and appropriate intervention. Below are the data fields:

  • Name: Name of the child for whom milestones are being recorded.

  • Age in months: Age in months at time of assessment.

  • Milestone name: Milestone being evaluated.

  • Assessment: Indicates the outcome of the milestone assessment for the child in question.

  • Milestone Assessment Month: Refers to the specific month in which the milestone was assessed for the child, which can be assessed as P - Present and A - Absent.

  • Starting month for milestone assessment: Indicates the starting month in which the specific milestone can be evaluated.

  • Deadline month for milestone assessment: Indicates the cut-off month up to which the specific milestone must be evaluated.

After creating the datasets containing examples of fictional children, filtering was conducted to identify only developmental milestones compatible with each child’s age group. Subsequently, another filtering step was performed with the objective of selecting exclusively the milestones that are present or absent for each child. The consistency of development milestones in relation to the child’s current and previous age group is also verified. These filtering aim to eliminate any noise or incorrect information that may be present in the data, ensuring that they do not negatively interfere with the inference and application of the rules.

Processing takes place to generate the knowledge base, with the facts and rules defined to be consulted via logical programming (Prolog). Prolog has advantages over complex problems through logical rules, facts and queries, which facilitate the clear expression of domain knowledge. Furthermore, Prolog employs a depth-first search pattern, exploring all possibilities until a solution is found, which is useful for problems with multiple alternatives.

The knowledge base is generated iteratively (for each child and child’s milestone) according to the scheme in Fig. 1.

Fig. 1.
figure 1

Knowledge Base Structure. (A) - Knowledge base model; (B) - Example knowledge base facts and rules.

Full size image

Subsequently, all rules are consulted for each child, generating a checklist: whether each milestone was met or not. If the list elements return true indications, then the SUITABLE classification is obtained. Otherwise, a new list is computed only for the missing landmarks. This new listing consists of which missing milestones indicate’alerts’ or’delays’ in the child’s development, as described in the Algorithm 1.

Algorithm 1
figure c

List of Missing Milestone

Full size image

If the list generated by the algorithm has only one element as’delayed’, then the classification is assigned as DELAYED for a given child, otherwise the child is classified as ALERT.

Fuzzy System Modeling. Subsequent to the execution of the first module, based on first-order logic, the fuzzy machines receive output data referring to the months in which the child reached the milestones of a certain subset, according to the Child Development Framework (i.e. green, pink, blue , yellow). Such data are submitted to fuzzification procedures in order to be represented in linguistic terms, through concepts that express the presence and/or absence of the landmark within a time interval (age group). Figure 2 shows the relationship between the fuzzified input variables (antecedents), their respective defuzzified outputs (consequent) and the meaning of the score for each membership function.

The result of the defuzzification process consists of a score from zero to one hundred, organized into relevance functions that represent the child’s performance linked to a certain set of milestones, shown in Fig. 2. At the end of the execution, the score obtained by the Fuzzy control system is converted into a linguistic variable that classifies the child’s performance. It is important to highlight that the fuzzy sets of “very high", “high", “low" and “very low" scores refer to the number of late milestones present (low score), as well as milestones reached before the start of the track. age established by the booklet (high score) (see Fig. 2).

Fig. 2.
figure 2

List of Antecedents and Consequents (SDIA’S Fuzzy inference machine).

Full size image

Regarding the inference rules, the following quantities were produced for each Fuzzy machine associated with the development milestone sets: 16 (Green); 12 (Pink); 14 (Blue); 9 (Yellow). Such quantities are justified by the nature of the problem, which does not require a high number of rules, thus eliminating the risk of redundancy, as well as deterioration of the capacity of the predictive model. Some samples of inference rules are presented in Algorithm 2.

Algorithm 2
figure d

Samples from Inference Rules applied to SDIA’s Fuzzy Inference Machines.

Full size image

After execution, the fuzzy module is capable of generating responses that indicate to the specialist professional the trend and status of child development in a given set of milestones, which allows it to contribute to a better understanding of the performance associated with the child’s evolution.

4 Results

To collect results, the hybrid artificial intelligence model was used to classify the child and developmental milestones based on data from 12 fictional children. In order to integrate the data from the hybrid model, the Fuzzy system will use as input the markers of months in which the child presented a certain milestone, as well as the monitoring signal generated by Prolog (i.e. ADEQUATE: AD, ALERT: AL, DELAY: DL). Table 3 presents the number of milestones achieved in a period of up to six months, as well as the score generated by the fuzzy system.

Table 3. Number of milestones reached within six months and score generated by the fuzzy system
Full size table

The children evaluated in  3 classified according to the score obtained in the score and described in intervals in the universe of discourse (see Fig. 2):

  • Pink, Blue and Yellow Fuzzy Machines:

    • Between 0 and 40 (VERY LOW and LOW - Trapezoidal): Presence of late milestones (below or above two milestones, respectively);

    • Between 20 and 80 (MEDIUM - Bell Function): Milestones achieved in the first months (over 50) or in the second half of the age group (under 50);

    • Between 60 and 100 (HIGH and VERY HIGH - Trapezoidal): Presence of milestones reached before the predicted age range (below or above two milestones, respectively);

  • Green Fuzzy Machine:

    • Between 0 and 80 (ATTENTION and ABSENCE OF MILESTONES - Trapezoidal): Set of functions that represent the number of milestones not reached by the child;

    • Between 80 to 100 (MELTERMS ACHIEVED - Trapezoidal): All milestones have been achieved;

After calculating the score, the system will classify the performance on each fuzzy machine based on the score and the intervals described above. At the end of the execution, the specialist will have a general overview of the child’s development, which is described by the code below:

  • Green Fuzzy Machine:

    • A: Normal Development;

    • B, from 1 to 4: Development alert - Delay in Basic Milestone detected. Level B1 represents one late milestone, while B4 represents all late milestones;

  • Blue, Pink and Yellow Fuzzy Machine:

    • A, from 1 to 2: Normal Development. Level A1 represents the milestone reached in the first half of the age group, while level A2 represents the milestone reached late;

    • B, from 1 to 2: Milestone delay detected. Level B1 represents one milestone late, while B2 represents two milestones late;

    • B, from 3 to 4: High number of Milestone delays detected. Level B3 represents three milestones late, while B2 represents all milestones late;

    • C, from 1 to 4: Milestone detected early (before the standard age range). Level C1 represents one milestone advanced, while C4 represents all milestones advanced;

Based on the established parameters, it becomes possible to map the current state of development. When applied to the sampling described in Table 3, we have the evolution profile of each child for each set of milestones. Table 4 presents the interpretation of the results according to the score obtained by the child in each inference machine.

Table 4. Development profile and interpretation of fuzzy outputs
Full size table

The results arising from this hybrid structure make it possible to establish a system capable of monitoring and evaluating progress in child development, providing a robust mechanism that addresses the complexity inherent in evaluating the growth and learning of children up to 6 months of age.

5 Final Considerations

In this study, the implementation of an artificial intelligence system is presented that integrates the logical reasoning capabilities of Prolog with the flexibility of fuzzy control systems to monitor child development. In this initial version, the development milestones of the first six months were addressed, which demonstrated satisfactory accuracy.

However, the need for additional refinements in fuzzy inference rules is evident. The accuracy and relevance of the system are intrinsically linked to the quality of these rules. Furthermore, expanding the model to include developmental milestones up to age 6 presents an opportunity to expand the scope and applicability of the system, ensuring its relevance and usefulness to a broader spectrum of child development.

For future work, it is suggested to integrate Generative AI into the fuzzy system to automatically predict and refine inference rules. This would increase the accuracy of the system by adapting it to emerging trends as the child grows. Furthermore, it is recommended to implement a graph model to relate development milestones and improve system analysis. In summary, the initial success of this system in monitoring child development indicates a promising path for future innovations in this field. Furthermore, it is also suggested a comparison with other similar systems present in the literature, in addition to the analysis of approaches such as non-monotonic reasoning and refutable reasoning, which can offer greater flexibility for solving the problem and developing new algorithms.