Impaired parental intersubjectivity frequently manifests as a parent's rigid misattributions of a child's unwelcome behavior to malevolence. For example, a mom may think her son's defiance to her requests means ill to her, “he did it to humiliate me.” When repeated misattributions of the child consolidate into a rigid belief, the parent may interpret all difficulties in parenting as a character flaw in the child, “he is mean.” Such problem is called “over-mentalizing,” i.e., the parent overly mentalizes the child into a generalization without relying on situational cues into circular reasoning “he defies me to humiliate me because he is mean.” Holding on to such a misbelief, the parent can develop a judgmental stance toward the child, which subsequently predicts chronic rejection, rage toward the child, parent-child bonding problems, and parental depressive moods (23). Furthermore, when parents habitually over-mentalize the child, they ignore situational, emotional, and behavioral cues in the “real-time” parent-child interactions that could otherwise serve as bottom-up data to rectify the parents' misbeliefs (24). Such obliviousness is called “under-coupling,” i.e., the parent is disengaged from observing how their physical or verbal actions (e.g., negative judgments or rejections) make their child feel and may “induce” the observed behaviors. Both “over-mentalizing” and “under-coupling” are undesired mental state manifestations and indicators of impaired parental intersubjectivity. When parents experience heightened parental stress, their defensive reactions (e.g., fight or flight) become sensitized, and “over-mentalizing” and “under-coupling” phenomena can worsen, further exacerbating impairment of parental intersubjectivity in a vicious cycle.

To mitigate parenting problems and reduce parenting stress, our team has developed MP, a group parenting intervention that fosters maternal intersubjectivity in clinical settings. For details on the intervention delivery, please see elsewhere (25). Impact on intersubjectivity is thought to be accomplished through (1) interpersonal, interactive exchanges with group peers to facilitate implicit imitations and explicit empathy-boosting exercises, (2) hands-on acquisition of knowledge regarding child's developmental needs to rectify developmental expectations and improve mothers' working models/mental representations of their child, (3) non-judgmental mindfulness practice to support regulation of own distress, which in turn inhibits mothers' defensive reactions to stress, and (4) enhancement of reflective capacity to build the awareness of self and other's lived experiences and needs (24, 26–28). Previously, we have found that MP reduces parenting stress (27), corrects developmentally-inappropriate, distorted working models/mental representations of their child (28), and modulates maternal brain responses to baby cry stimuli as a function of parenting stress (29). Based on this work, we postulate that MP will reverse both maternal intersubjectivity problems, “over-mentalizing” and “under-coupling” (28), which in turn will reduce maternal parenting stress.

The social neuroscience literature suggests that the recognition and attribution of goals and intentions of another person's behaviors is primarily supported by three distinct but inter-related neural systems, namely mirroring system, mentalizing system, and salience network, described below (30–32). The mirroring system becomes active when an agent performs an action or perceives another agent's similar action (33). The co-localization of activities related to perception and action in the brain affords an observer's automatic recognition of the immediate goal of the other agent's actions. This system involves the posterior inferior frontal gyrus (pIFG), dorsal and ventral premotor cortex (dPMC and vPMC), supplemental motor area (SMA), inferior parietal lobule (IPL), superior parietal lobule (SPL), intraparietal sulcus (IPS), superior temporal gyrus (STG), and pericentral cortex (34–39).

The mentalizing system becomes active when a person is attributing mental states to others and this system involves the precuneus/posterior cingulate cortex (PrC/PCC), dorsal, middle, and ventral medial prefrontal cortices (dmPFC, mmPFC, and vmPFC, respectively), posterior temporal sulcus (pSTS), temporal pole, and temporoparietal junction (TPJ) (32). A meta-analysis suggests that the dmPFC, mmPFC, vmPFC, and PrC/PCC form a loop to generate narrative thoughts related to affective representations of self and other (40). In this loop, interpersonal scripts (autobiographical stories) are generated when the PrC/PCC, as a thought generator (41), connects affective potentials stored in the vmPFC (42) to regions that serve as a proximal-object sketchpad that represents the self (in mmPFC) (43, 44) or a distal-object sketchpad that represents another person (in dmPFC) (43). The dmPFC-dependent functional connectivity preferentially participates in mentalization in verbal, but not in visual, modality, while the TPJ-dependent functional connectivity participates in both modalities (45). Thus, the dmPFC represents others' enduring attributes (a generic image of other's identity) without differentiating self and other's perspectives (40, 46). In contrast, the TPJ represents other's inner thoughts that are different from one's own perspective, with self-other distinction (47) and mediates inferences about others, such as their transient goals, desires and beliefs (48). Moreover, the anterior part of TPJ is involved in joint attention, which requires spatial representation of other's attentional direction (30).

The salience network includes dorsal ACC, posterior ventral MCC, bilateral anterior insula cortices (IC), and subcortical regions such as PAG, hypothalamus, thalamus, midbrain, striatum, and extended amygdala (49). This network detects internal and external events that are personally meaningful (50) and interacts with the mentalizing system to respond to attachment figures (51). Indeed, the salience network largely overlaps with a maternal caregiving system that regulates parenting behaviors, including the amygdala, IC, and two motivational sub-systems—one for affiliative motivations that include the hypothalamus, ventral tegmental area (VTA), nucleus accumbens (NAc), and ventral pallidum (VP) and the other for defensive (fight or flight) motivation mediated by the periaqueductal gray (PAG) (52–54). Notably, many of these regions (e.g., the amygdala, PAG, and NAc) are sensitive to signed prediction errors of reward or punishment with reference to preceding baselines, i.e., activated when detecting a greater-than-expected level of salience (positive prediction errors of reward or punishment, e.g., the presence of unexpected salience) and deactivated when detecting a less-than-expected level of salience (negative prediction errors of reward or punishment, e.g., the omission of expected salience). For examples, the NAc is sensitive to signed prediction errors of reward (55, 56); the amygdala is sensitive to signed prediction errors of reward (e.g., desirable liquid) and/or punishment (e.g., undesirable air-puff) (57); besides, the amygdala is also sensitive to the signed prediction errors in aversive stimuli, e.g., activated when detecting the presence of unexpected foot shock and deactivated when detecting the omission of expected foot shock (58). Notably, consistent with the notion that NAc and PAG served as opponent motivations of reward-seeking and defense (flight-or-flight) respectively, the NAc and PAG responded in opposite manners to aversive prediction errors, as unexpected pain not only deactivated the NAc (a negative prediction error of reward as if the unexpected pain was equivalent to the omission of reward), but also activated PAG (a positive prediction error of punishment as if the unexpected pain was equivalent to the presence of unexpected punishment) (59). As described later, the contrast of mirroring the child's joyful vs. distressed expressions in our experimental task is computed to index the sensitivity of signed prediction errors specific to maternal mirroring of own child's emotions.

The mirroring system largely overlaps with the frontoparietal network (60); the mentalizing system largely overlaps with the default-mode network that is more active during resting states (61) and mind wandering (62), as compared to states of actively paying attention to the environments. Spontaneous activities in the default-mode network are often anti-correlated with those in the frontoparietal network (63). Thus, we postulate that empathic mirroring of others encompasses bottom-up perception-action coupling between two agents, which can potentially activate the mirroring system and automatically deactivate the mentalizing system, as compared to (non-mirroring) observing others.

Moreover, as virtually all cognitive processes depend on the functional connectivity among participating brain networks (64), the functional connectivity among the three aforementioned brain systems, i.e., mirroring system, mentalizing system, and salience network, are key to intersubjectivity (45). Indeed, the capacity for intersubjectivity seems to depend on the functional connectivity between the dmPFC (in the mentalizing system) and the inferior frontal gyrus (in the mirroring system) (65). Notably, the functional connectivity between stress-dependent brain regions (which include the salience network) and the child-representing regions, i.e., dmPFC, may underlie the stress-potentiation of the “over-mentalizing” problem. It is through functional connectivity that the salience network may switch up or down the activity in the frontoparietal network and the default-mode network alternately (66). These results underscore the roles of dmPFC-dependent functional connectivity in representing the child in maternal intersubjectivity. Moreover, the pain-related prediction error signals in the PAG are functionally connected to the dmPFC (59). Thus, we postulate that the functional connectivity between the dmPFC and PAG should reflect the extent to which maternal defensive motivation can influence the representation of the child. In other words, we postulate that the dmPFC-PAG functional connectivity should modulate the maternal mirroring of the child as a function of parenting stress.

To provide a theoretical relationship between interpersonal stress and the “over-mentalizing” and “under-coupling” problems at an abstract level of analysis, we postulate a dyadic active inference framework in a two-agent system, which will be elaborated in section Abstract Level of Analysis—Toward an Overarching Framework for Research on Intersubjectivity. Our framework is inspired by Karl Friston's Free Energy Principle (67, 68) and its application to stress (69). In brief, this framework postulates that in a two-agent system, stress ensues in a dyad when an agent's working model of the other agent in the system results in excessive prediction errors in a way that threatens the agent, and the stress worsens when the agent's preconceived working model of the other agent defies, rather than accommodates, the prediction errors. On the basis of this theoretical framework, we are led to postulate that when a mother shows symptoms of impaired intersubjectivity during mother-child interactions, she is at risk for excessive stress, and that her capacity to empathically mirror the child's actions and feelings may be compromised, reducing her sensitivity to the child's feelings, especially when the child's expressions are incongruent to the mother's preconceived working model/mental representation of her child. Thus, when she is stressed and/or in a negative mood, her mirroring of the child's joyful expressions may be diminished (i.e., stress-potentiated “under-coupling”), while her defensive reactions to the child's distressed expressions due to her preconceived working model may be exacerbated (i.e., stress-potentiated “over-mentalizing”). From prior work, we know that MP changes mothers' mental representations/working models toward less distorted/rigid/negative perceptions (28).

As the brain bases for “over-mentalizing” and “under-coupling” problems may be inferred through various experimental tasks in neuropsychiatric disorders (70), in the present study we employed a face and affect imitation task, namely Child Face Mirroring Task (CFMT) in the fMRI setting, which has been substantially modified from a previously published task (71). The CFMT involves pictorial displays of children's facial expressions, sorted in three independent factors, Child's Identity (Own Child and Other's Child), Emotions (Joy, Distressed, Ambiguous, and Neutral), and Task (Join, Observe, and React). In a full factorial design, each of the two children's pictures displayed four kinds of emotional expressions (Emotions), and all these pictures are repeated in three distinct conditions (Tasks): a face/affect mirroring condition (Join) and a non-mirroring control condition (Observe) to evoke strong and weak mother-child coupling, respectively, and, additionally, a React condition in which mothers respond to child faces as they normally would, to examine whether MP changes mothers' voluntary (un-instructed) responding. Results from two studies are reported here. In Study 1, in a sample of healthy mothers (n = 45), we examined the main effects of CFMT. In Study 2, we used CFMT in an randomized controlled intervention study where mother either receive the MP intervention (n = 19) or are in Control condition (n = 17), and measured maternal parenting stress at both pre- and post-treatment time points (T1 and T2) to identify potential neural mediators of MP effects on parenting stress.

Using CFMT, we computed a family of contrasts, namely Maternal Mirroring Response (MMR), to examine neural underpinning of own-child-specific maternal intersubjectivity. As these contrasts will be included in our predictions, we need to describe them before we prescribe the predictions. To isolate maternal neural responses in child-specific empathic mirroring across all emotions, we construed a MMR(all) contrast, i.e., Join[Own vs. Other's child's all emotions] vs. Observe[Own vs. Other's child's all emotions]. We also examine the contrast of positive vs. negative emotion in MMR, namely MMR(j-d), i.e., Join[Own vs. Other Child's Joyful vs. Distressed] vs. Observe[Own vs. Other Child's Joyful vs. Distressed]. The MMR(j-d) contrast approximately indicate the range of signed prediction errors, i.e., the range of MMR = MMR(j) - MMR(d), assuming that mirroring own child's joyful expression MMR(j) and distressed expression MMR(d) should elicit the maximum and minimum of prediction errors respectively in the brain regions that are sensitive to signed prediction errors. When these regions' sensitivity to signed prediction errors is diminished, e.g., MMR(j) is not different from MMR(d), then the range of MMR(j-d) should be no different from zero. The reasons for examining MMR(j-d) include: (1) as described above, the amygdala, NAc, and PAG are sensitive to signed prediction errors in emotional salience (reward or punishment), we postulate that these regions' MMR responses to positive (joyful) and negative (distressed) expressions may differ in the directions, e.g., relatively activated in MMR(joy) and deactivated in MMR(dis) for the amygdala and NAc, and vice versa for PAG; (2) the child's positive vs. negative facial expressions have been found to differentially activate maternal amygdala (72) as a function of unresolved stress (73), thus the maternal amygdala's sensitivity to the child's emotion during empathic mirroring may vary as a function of maternal stress. Taken together, the literature suggests that the maternal amygdala should be sensitive to MMR(j-d) and parenting stress may diminish the MMR(j-d) in the amygdala.

Based on the literature discussed above, we hypothesized that MP can reduce parenting stress by improving the mothers' working models of the child toward more flexible and positive perceptions, which can in turn improve maternal empathic mirroring of the child's joyful expressions (treating “under-coupling”) and can prevent the mothers' defensive reactions from coloring their mental representation of the child (treating “over-mentalizing”) during empathic mirroring. This hypothesis would be translated to the following group (MP vs. Control) by time (T1 vs. T2) interaction effects in the present study: We predict that, from T1 to T2, MP (vs. Control) will (1) reduce parenting stress measured with parenting stress index (PSI); MP (vs. Control) will rectify the “over-mentalizing” problem by (2) decreasing the mentalizing system activities during own-child-specific voluntary responding (React to Own vs. Other's Child); MP (vs. Control) will rectify the “under-coupling” problem by (3) increasing MMR(all) (own-child-specific empathic mirroring) in the mirroring system and by (4) increasing MMR(j-d) in the amygdala that mediates signed prediction errors of emotional salience. Because parenting stress can potentiate the “over-mentalizing” and “under-coupling” problems, we also predict that (5) the reduction in parenting stress will be associated with the reduction of the “under-coupling” problem, which can manifest as the association between the reduced parenting stress and increasing sensitivity to the signed prediction errors in the amygdala's MMR(j-d); (6) the reduction in parenting stress will be associated with the reduction of defensive “over-mentalizing,” which can manifest as the association between the reduced parenting stress and decreasing MMR(all)-dependent functional connectivity between the dmPFC (the sketchpad for child representation) and the PAG (the signals for defensive, fight-or-flight motivation). To summarize these predicted effects succinctly, we used non-parametric mediation analyses to identify potential neural mediators of MP treatment effects on parenting stress.

The research reported in the current study was approved by the Institutional Review Board (IRB) at the University of Michigan, Ann Arbor, Michigan, USA. Informed consent from all participants was obtained. All research was performed in accordance with relevant IRB guidelines/regulations.

All participants were recruited from low-income community clinics, primary care clinics, and/or community mental health centers. In Study 1, we examined brain responses during CFMT in a sample of healthy, unmedicated participants who underwent the CFMT (see below) for the first time (n = 45, age M = 31.78, SD = 7.62, child age M = 2.61, SD = 2.05). As MP's efficacy in reducing parenting stress has been established previously (27, 29), we conducted Study 2 to examine MP effects on intersubjectivity-dependent maternal brain responses and how these responses are associated with reduction in parenting stress. In Study 2, participants (n = 36) were randomly assigned to either MP treatment group (n = 19) or Control group (n = 17) and underwent the CFMT before (T1) and after (T2) MP or Control conditions, with about 14 weeks between scans. The participants in MP and Control groups differed slightly in their age [MP: M = 27.84, s.e. = 1.71; Control: M = 33.35, s.e. = 1.81, F_{(1, 34)} = 4.92, MSerror = 55.42, p = 0.033], but there was no group difference in the child age [MP: M = 2.25, s.e. = 0.40; Control: M = 3.09, s.e. = 0.42, F_{(1, 34)} = 2.08, MSerror = 3.06, p = 0.16] and number of offspring [MP: M = 1.63, s.e. = 0.19; Control: M = 1.65, s.e. = 0.20, F_{(1, 34)} = 0.003, MSerror = 0.66, p = 0.96]. There were three and five participants in MP and control groups, respectively, who were medicated with steady dosing anti-depressants across the study period. Nevertheless, we expected that the potential effects of medication would be canceled out for the following reasons: (1) medicated cases were in the minority and similarly distributed across MP and control groups (Chi-square Z = 0.963, p = 0.33), and (2) the repeated measures design controlled for the heterogeneity in medication status as participants are compared to their own baseline. As described further in the Supplementary Materials, removing medicated participants did not change results.

For the illustration of the task design, see Figure 1. In CFMT, participants were presented repeatedly with the same pictures of Own and Other Child in three conditions (Tasks), namely Observe, React, and Join. By design, the Observe Task should elicit the participant's unresponsive observation of face-like visual objects (i.e., “look-at-it,” a weak coupling condition), React should elicit the participant's usual, voluntary responses to the presented child, and Join should elicit the participant's empathic mirroring of the presented child (i.e., “empathize-you,” a strong coupling condition). The task instructions were presented to study participants as follows.

Observe: “You should simply observe the face on the screen. You should NOT make any face or generate any emotion. That is, BE an OBJECTIVE viewer of the faces. DO NOT FOLLOW any feelings depicted or caused by the face.”

React: “You should react to the emotion and expression of the child on the screen. You should imagine that you are the caregiver of the child on the screen. That is, you are REACTING to the emotions of the child on the screen as you normally would in your home.”

Join: “You should Join your own emotion with that of the emotion and expression of the child on the screen. You should empathize with the emotion depicted on the screen. That is, you are JOINING in the emotions of the child on the screen, with your OWN emotions.”

The three Tasks were presented block-by-block in a pseudo-random order. There were four pictures of a single child (one picture each for neutral, ambiguous, Distressed, and joyful expression), presented consecutively in a pseudo-random order, in each block (4 s each picture, 16 s per block). There were 4 blocks per Task for each of the Own and Other Child, with 10-s resting intervals between the blocks. To ensure the participants' wakefulness and readiness for the task, before each block, a single-word cue (“Observe,” “React,” or “Join”) was presented on the screen and participants pressed a button to indicate as soon as they were ready to perform the Task as instructed, without knowing which child's pictures would be presented. The reaction time in pressing the button was defined as Cue Period, reflecting the time required for a participant to be ready to perform the following task. The statistical analysis of the reaction time in Cue period is reported in Supplementary Materials.

The participants provided all their child's pictures used in the study. The pictures of children unknown to the participants (Other's Child) were drawn from the in-house inventory. The lab staff standardized the stimuli qualities based on specific expressions (neutral, ambiguous, distressed, and joyful). We included these four kinds of expression following pioneering work in the field of parental neuroscience (71). Ratings of the child emotional expression images using Manikin Self-Assessment Scale (74) by four independent female raters confirmed the validity of the valence of the stimuli and the valence and arousal level were matched between the Own and Other Child's pictures, as described in Supplementary Materials and Supplementary Figure 1.

Before each scan, the participants practiced CFMT to ensure their comprehension of the task and minimized effects due to stimuli novelty or learning. In each MRI scan, the participant was positioned in a supine orientation with her head positioned in a head coil. Visual stimuli were presented with E-Prime (PST, Inc., Pittsburgh, PA), via a goggle system and Nordic NeuroLab audio system. Behavioral responses were recorded by a button glove attached to the participant's right hand and linked to the E-Prime system. All fMRI scans were performed with a 3.0 Tesla Philips magnetic resonance imaging scanner using a standard 8-channel radiofrequency SENSE head coil with the following acquisition parameters: (1) A high-resolution T1 scan was acquired to provide precise anatomical localization (TR of 9.8 ms, TE = 459 ms, FA = 8°, FOV of 256 mm, slice thickness of 1.0 mm, 180 slices with 288 × 288 matrix per slice). (2) Two runs of T2^*-weighted EPI sequence with BOLD (blood oxygenation level dependent) contrast (190 frames per run, TR = 2,000 ms, TE = 30 ms, FA = 90°, FOV = 220 mm, 42 contiguous axial slices, slice thickness = 2.8 mm with 64 × 64 matrix per slice, voxel size = 3.44 × 3.44 × 2.8 mm³) were acquired for whole-brain fMRI BOLD signal measures during the experimental task.

For both Study 1 and 2, MRI data were pre-processed and analyzed using statistical parametric mapping software (SPM8; Welcome Department of Imaging Neuroscience, London UK). Five images at the beginning of each fMRI run were discarded to account for magnetic equilibrium. Slice timing correction was performed using a middle slice as a reference (slice 21). After slice time correction, images within each run were realigned to the mean image of the first run to correct for movement. Realigned functional images and structural image were spatially normalized using DARTEL method in SPM8. The normalized functional images were re-sliced to 2 × 2 × 2 mm voxels. Images were then spatially smoothed using a Gaussian filter with a full-width half-maximum value of 8 mm. All the images in the analyses and the figures are in neurological convention, with the left hemisphere presented at the left of an axial image.

We tested the group differences in demographic variables and the MP treatment effects on parenting stress (as indexed by PSI) in GLMs, using SPSS v.24 (IBM Corp. Armonk NY). We also performed the non-parametric mediation analysis based on the bootstrapping of 5,000 times of re-sampling, with a covariate of Own Child's age, using the macro of PROCESS (86) in SPSS v.24 (IBM Corp. Armonk NY). Due to the space limit, the results of these non-MRI analyses are described in Supplementary Materials and Supplementary Figure 1 (Independent raters' rating on the stimuli), Supplementary Figure 2 (MP effects on PSI), Supplementary Figure 3 (Cue period of CFMT), and Supplementary Figure 4 (robustness check after removing medicated participants in Study 2).

We first report the results of primary main effects in CFMT. The main effects of Tasks (Observe, React, and Join vs. Rest) and the pairwise planned contrasts (React vs. Observe, Join vs. Observe, and Join vs. React), pooling across both children, are summarized in Table 1 and Figure 2, with the key brain regions depicted in Figure 3. As expected, all three Tasks activated face-related processing in visual cortex and fusiform face area (FFA). Interestingly, the neural responses in some of these visual processing areas were attenuated in both Join vs. Observe and React vs. Observe contrasts. Conversely, both Join vs. Observe and React vs. Observe contrasts activated brain regions involved in the mirroring system (32), including pericentral, insular, frontoparietal cortices, and thalamus, and the salience network (49), including striatum, and amygdala.

As summarized in Table 2 and Supplementary Figure 5, the brain regions that were conjunctively implicated in both Join vs. Observe and Join vs. React contrasts included the bilateral pericentral cortices and left inferior parietal lobule (IPL), which were more activated in Join than the other two Tasks, and the occipital and lingual cortices, right hippocampus, and the dmPFC, which were less activated in Join than the other Tasks (also depicted in Figure 3D).

The main effects of Child (Own vs. Other Child) are summarized in Supplementary Table 1. The neural responses in the occipital, precuneus, angular gyrus, and FIO cortices were greater in Own than Other Child. These regions are largely involved in autobiographical memory, thus consistent with their roles in the mentalizing system.

Since there were some Task-by-Child-interaction effects, described below, we examined the simple main effects of Own vs. Other Child in each Task separately (Supplementary Table 1). For Observe, we found that the Own vs. Other Child in this Task elicited differential neural responses in the visual face processing areas (FFA) and autobiographical memory-related regions (i.e., FIO, temporal poles and hippocampus), and cognitive regulatory regions (right dorsolateral prefrontal cortex (dlPFC) and supplemental motor area (SMA), which were more active in the main effects of Join (>Observe) and React (>Observe), indicating that the mothers automatically engaged the Own Child with more autobiographical and interactive responses than they did in Observe of Other Child, despite that the task instruction of Observe explicitly discouraged such active child-oriented responses. For React, the Own vs. Other Child elicited differential responses in the subcortical regions, including the thalamus, hypothalamus, striatum, hippocampus, and midbrain, suggesting the mothers responded to Own Child with greater maternal motivation than they did to Other Child. For Join, there were no Own vs. Other Child differences in any regions.

The planned tests related to Task-by-Child interaction effects [including MMR(all) and MMR(j-d)] are summarized in Supplementary Table 2. For MMR(all), we found that the precuneus and fusiform gyrus showed greater Own > Other differential responses in Observe than in Join (Observe > Join)—which is an inverse MMR(all)—suggesting that the Join, as compared to Observe, reduced the face processing, mediated by the fusiform gyrus (87), and narrative thinking processing, mediated by the precuneus (41), related to Own Child. We also found that the midbrain, striatum/extended amygdala, and hypothalamus showed greater Own > Other differential responses in React than in Join—suggesting that the own-child-specific maternal motivation responses were stronger in the React than Join. For MMR(j-d), we found that the left amygdala was associated with MMR(j-d) (MNI coordinates: [−26, 2, −24], 15 voxels, Z = 3.17, p = 0.021 s.v.c.).

We predicted MP treatment effects on parenting stress, maternal voluntary mirroring (probed in the React Condition) and maternal mirroring responses [MMR(all) and MMR(j-d)]. For parenting stress, we found that MP, relative to Control, showed lower PSI total scores at T2 (see Supplementary Materials). We examined MP Treatment effects by testing Group-by-Time interaction effects on three own-child-specific contrasts, i.e., React to Own vs. Other Child, MMR(all), and MMR(j-d).

For React to Own vs. Other Child, we found that, from T1 to T2, MP, relative to Control, decreased React to Own (vs. Other) Child responses in the dmPFC ([−8, 54, 14], 887 voxels, Z = 3.08, p = 0.009, whole brain cluster-level FWE corrected, Figure 4).

As the amygdala mediated MMR(j-d) in the reference sample (Figure 5A), there were several Group-by-Time interaction effects on the amygdala as follows. We found that MP, relative to Control, increased the MMR(j-d) in the left amygdala ([−24, −2, −18], 24 voxels, Z = 2.87, p = 0.046 s.v.c., Figure 5B), in which the differential response to own child's Joy expression was increased in MP, but decreased in Control, mothers.

These results suggested that MP mothers, relative to Control, developed stronger capacity not only to activate the left amygdala in response to own child's joyful faces when they were instructed to mirror the children's emotions in the Join condition, but also to inhibit the own-child-specific neural responses in the dmPFC (Figure 4) during their voluntary mirroring responses to their own child in the React condition.

For MMR(all), from T1 to T2, MP, relative to Control, increased the MMR(all) in the left frontoparietal regions including the parietal/postcentral ([−56, −26, 42], 357 voxels, Z = 4.22, p = 0.001, whole brain cluster-level FWE corrected) and dorsolateral prefrontal cortex ([−56, 16, 28], 706 voxels, Z = 3.48, p = 0.001, whole brain cluster-level FWE corrected, Figure 6A), midbrain ([10, −20, −4], 124 voxels, Z = 3.55, p = 0.049 s.v.c., Figure 6B), left NAc ([−8, 16, −12], 26 voxels, Z = 3.37, p = 0.013 s.v.c., Figure 6C), left amygdala ([−28, 2, −22], 12 voxels, Z = 3.18, p = 0.021 s.v.c., Figure 6D), and, marginally, right amygdala ([24, 0, −16], 18 voxels, Z = 2.85, p = 0.057 s.v.c., Figure 6E).

To examine the results in elementary conditions, such as specific tasks and emotions, we unpacked the elements involved in the Figure 5B in the Supplementary Figure 6. Similarly, we also unpacked the elements involved in the Figure 6F in the Supplementary Figure 7.

We hypothesized that MP can reduce parenting stress by improving the mothers' working models of the child, which in turn improve maternal empathic mirroring of the child's joyful expressions (reversal of “under-coupling”) and prevent mothers' defensive reactions from shaping their mental representation of their child (reversal of “over-mentalizing”) during empathic mirroring. The hypothesis was supported by the results in the following group-by-time interaction effects during the CFMT: We found that MP (vs. Control), from T1 to T2, (1) reduced parenting stress (Supplementary Figure 2), (2) decreased the dmPFC (in the mentalizing system) activities during own-child-specific voluntary responding (React to Own vs. Other's Child), suggesting that MP rectified the “over-mentalizing” problem (Figure 4), (3) increased MMR(all) (own-child-specific empathic mirroring) in the mirroring system (Figure 6), and (4) the amygdala's MMR(j-d), i.e., the sensitivity to the prediction errors (Figure 5), suggesting that MP rectified the “under-coupling” problem. The results also supported our predictions that, from T1 to T2, MP (vs. Control) (5) reversed the stress-potentiated “under-coupling” problem, suggested by the association between the increasing sensitivity to signed prediction errors in the amygdala's MMR(j-d) and the decreasing parenting stress index (PSI) (Figure 9) and (6) reversed the stress-potentiated over-mentalizing problem, suggested by the association between the decreasing MMR(all)-dependent dmPFC-PAG functional connectivity and the decreasing PSI (Figure 8). We also identified three potential brain mediators of the MP treatment effects on reducing parenting stress: (1) the T1-to-T2 increases in the MMR(all) of the left STG, (2) the T1-to-T2 decreases in the MMR(all)-dependent psychological-physiological interaction (PPI) between the dmPFC and PAG, and (3) that the T1-to-T2 increases in the MMR(j-d) of the left amygdala. The results of these potential mediators will be discussed later.

In Study 1, the results in the contrast between strong coupling (Join) and weak coupling (Observe) conditions is highly consistent with the predictions deduced from our novel dyadic active inference framework. Specifically, the Join > Observe contrast primarily activated the mirroring system, along with the salience network, including the SMA, pericentral cortex, inferior parietal lobule (IPL), insula, thalamus, striatum, and left amygdala. Conversely, the Join > Observe primarily deactivated the mentalizing system, including the dmPFC, precuneus/posterior middle cingulate cortex, parahippocampal gyrus/hippocampus, and OFC, along with the visual cortex. Furthermore, some of these Join vs. Observe results overlapped with the Join vs. React results. Specifically, the strong coupling condition of Join (vs. both React and Observe) activated the bilateral pericentral cortex and left IPL, but deactivated the dmPFC, primary and secondary visual cortices, and right hippocampus.

According to the affect-object active inference model (40), the dmPFC may mediate the mentalization of others (as a distal-object sketchpad to hold affective active inference of a counterpart), and it has been found that the dmPFC mediated mentalization based on a self-centered, rather than other-centered, perspective (46). The down-regulation of the dmPFC responses during the strong coupling condition (Join) in the healthy mothers in Study 1 probably help preserve their maternal intersubjectivity by preventing the over-mentalizing problem, which may manifest as perspective mistaking that can happen when one overly relies on preconceived beliefs (88). In short, it is probably necessary to suspend (temporarily down-regulate) the prior-driven dmPFC to avoid the over-mentalizing problem and achieve a higher level of intersubjectivity in a strong coupling condition.

The dmPFC has been known to be sensitive to repeated stress (89, 90) and postpartum depression (91). In accord, we previously found that, when listening to own baby's crying, the maternal dmPFC response (92) and its functional connectivity with anxiety-dependent extended amygdala (93) increased with maternal stress-related symptoms. The present study suggested a new insight into the roles of dmPFC in stress resilience, i.e., the dmPFC mediated maternal preconceived beliefs of the child as part of the mentalizing system, which should be temporarily suspended when the mothers relied on the mirroring system to empathically mirror the child. Moreover, MP enhanced the maternal capacity to down-regulate the dmPFC voluntarily while responding to own child and probably reduced parenting stress by diminishing the influences of PAG-dependent defensive/aggressive motivation signals on the dmPFC-dependent (preconceived) representation of the child. In other words, interpersonal stress can be reduced if defensive signals from the PAG are prevented from influencing the dmPFC, otherwise it would cause the defensive over-mentalizing problem that tends to increase stress.

With regard to the amygdala, we found that, in Study 1, (a) the left amygdala was activated in Join vs. Observe and (b) the left amygdala was sensitive to MMR(j-d); in Study 2, (c) from T1 to T2, the left amygdala's MMR(joy) (Join vs. Observe of Own vs. Other Child's Joyful expression) increased in MP, relative to Control, (d) from T1 to T2, the bilateral amygdala (and other regions in the maternal motivation and mirroring component) increased their MMR(all) responses in MP, relative to Control, and (e) T1-to-T2 increases in the left amygdala MMR(j-d) responses mediated the MP effects on reducing parenting stress.

The constellation of amygdala-related results provided more nuanced understanding of the amygdala's role in maternal behaviors, in accordance with the literature documenting the roles of amygdala in parental synchrony in interactions with the infant (94), empathy for the own child (71), positive feelings and attachment to the infant (95), and autobiographical recall of positive and negative emotion cues (96).

Consistent with the roles of NAc and PAG in maternal affiliative and defensive motivations, respectively (52, 53, 78) and their roles in signed prediction errors of reward (55, 56) and pain (59), respectively, we found that these two regions were related to the T1-to-T2 changes in parenting stress in opposite directions. While the T1-to-T2 changes in parenting stress were negatively associated with the NAc's MMR(j-d) and MMR(all)-dependent PPI with the dmPFC, it was positively associated with the PAG's. Consistent with the affect-object active inference model (40), these results highlights the role of dmPFC as a distal-object sketchpad in representing the child and the “coloring” of the representation with affiliative and defensive affective potentials, forming “affect-objects,” by its connectivity with NAc and PAG (59), respectively. So, this suggests that the role of affect-object generation during empathic mirroring in parenting stress, i.e., mirroring the child with affiliative or defensive affective potentials can decrease or increase parenting stress, respectively.

We also found that the T1-to-T2 reduction in parenting stress was associated with the concomitant increases in the MMR(all) in the left STG, right STG/Insula, cerebellum, and hypothalamus. Interestingly, the first three regions were related to music-entrained movement coherences in professional dancers (97), suggesting that increasing coherence in empathic mirroring may be related to parenting stress reduction. In a cross-culture study, these brain regions were commonly activated when mothers from different cultures listened to their own baby's cry (98). Consistent with the result that the left STG mediated the MP effects on reducing parenting stress in the present study, we have reported that the T1-to-T2 parenting stress reduction was associated with the concomitant increases in the functional connectivity between the left STG and amygdala, when the mothers responded to own baby's crying (29). Maternal STG responses to own vs. other's infant cry were associated with child-oriented caring thoughts and indirectly with infant development (99). Taken together, these results implicated that parenting stress reduction may depend on increasing the coherence in maternal empathic mirroring of the child, which is potentially mediated by the amygdala-STG neurocircuits, as part of the mirroring system.

We also found that the left prefrontal cortex was activated in Join vs. Rest (Figure 2C), in accord with a recent hyper-scanning study that reported increasing maternal parenting stress was also associated with the differences between mother and child's left prefrontal cortex responses when the dyads watched videos together (100). Considering that the left prefrontal cortex is part of the mirroring system (32), these results suggested that parenting stress may influence the mother-child coupling via the left prefrontal cortex.

Several limitations of the present study should be noted. First, although Study 1 established the intended effects of CFMT with whole brain correction in a relatively large sample (n = 45), the sample sizes of MP and Control groups in Study 2 were modest and thus the results should be considered preliminary and warrant future study. Second, there was heterogeneity in medication use in Study 2. Nevertheless, this heterogeneity would cause little confounding because not only it was partially controlled in the repeated measurement effects based on each participant's own baseline, but also the medicated participants were in minority and evenly distributed between the groups. As reported in Supplementary Materials, removing all medicated participants in statistical analysis did not qualitatively alter the results reported above. Third, we did not incorporate measurements that are directly linked to the “over-mentalizing” and “under-coupling” problems in the parenting context. Nevertheless, the effectiveness of MP's improvement on the symptoms of “over-mentalizing” and “under-coupling” has been documented (28) and thus the reported Time-by-Group interaction results should be closely related to the correction of “over-mentalizing” and “under-coupling” problems. We will examine the associations between the neuroimaging data and these variables in the future.

Bayesian active inference (also known as predictive coding) is a computationally powerful framework, as its variants not only can account for perception, cognition, emotions, and consciousness in humans and animals (35, 40, 67, 69, 102–108), but also biologic evolution (109) and even artificial intelligence (110, 111).

According to FEP (67, 68), an agent's predictive-coding engine can be heuristically modeled in a hierarchical network, which contains four nodes (E, S, A, and M) in three levels: E is Event from environments at the bottom, S is Sensation and A is Action at the intermediate, and M is the internal prior Model at the top level (Figure 12A). When an event E causes S to generate afferent data, S causes M to predict what the event means based on stored prior causal models, and M in turn causes A to respond to the event, and then the differences between S (the afferent data) and A (the efferent prediction) are computed, serving as prediction errors in feedback to update the priors in M. Because there is no direct contact between M and E nodes, the engine depends on the prediction errors resulting from the interactions between the agent's S and A to infer the events in E. The interactions between an agent's S and A and events (E) update the internal model M iteratively, until the prediction errors are minimized, and M is thus optimized.

The notion of human as a social active inference engine has emerged in the recent literature (35, 40, 106, 108). As social interactions lie at the core of intersubjectivity, single-agent active inference framework is simply inadequate to account for intersubjectivity. Thus, we propose a novel dyadic active inference framework to model intersubjectivity (Figure 12B). In this dyadic framework, in a two-agent coupled system wherein Agent 1 (say, Mom) and Agent 2 (say, Son) are strongly coupled such that one agent's action (A) predominantly causes the other's sensation (S) and vice versa, i.e., A_Mom ≈ S_Son and A_Son ≈ S_Mom, each agent's internal model (M) will serve as the other's events E, i.e., M_Son ≈ E_Mom and M_Mom ≈ E_Son. When Mom's internal model (her working model of the child) approximate Son's (his working model of the mother), M_Mom ≈ M_Son, she achieves intersubjectivity and minimizes her prediction errors in the dyadic system.

One question arises that if M_Mom ≈ M_Son, then the mother's working model (M_Mom) will be as helpless as the son's (M_Son) when the son struggles in distress. This would not be the case if the mother would possess more knowledge or wisdom, i.e., if her working model could access more repertoires or strategies that the son's does not have. It is important to note that the presence or absence of M_Mom ≈ M_Son as a state is transactional, not permanent. Therefore, after the mother achieves the state of M_Mom ≈ M_Son, she can access additional resources, repertoires, or strategies and then teach the son to expand his working model to solve his issue at hand. Conversely, without first achieving the state of M_Mom ≈ M_Son, the mother may fail to address what the son needs or to teach him any new strategies effectively because she may have misunderstood what the son actually needs in that current moment.

Our dyadic framework can make sense of why intersubjectivity can automatically minimize stress in a two-agent coupled system. Recently, stress has been re-defined as uncontrollable prediction errors (excessive free energy) that threatens the agent as a Bayesian active inference engine (69). Thus, the minimization of prediction errors is equivalent to the minimization of stress. When two or more agents are coupled as a relational whole, if one agent merely projects one's own beliefs about another agent's perception, action, and intention—without relying on data from ongoing dyadic interactions—the prediction errors will tend to increase, as exemplified in perspective mistaking (88).

How can imitation facilitate intersubjectivity? In the two-agent system, when the mother imitates the child's action (e.g., smile), their actions are similar and their perceptions are also similar (e.g., joy). By virtue of such reciprocal similarity, the dyad can better predict each other's covert working models underlying their actions and perceptions. Thus, imitation can reduce prediction errors in predicting each other's actions and feelings, which may in turn increase the similarity between their covert working models underlying those actions and feelings, thereby facilitating intersubjectivity. However, when the mother's preconceived working model of the child is fixated in negative mood under stress (excessive prediction errors), her capacity to utilize prediction errors to update her working model of the child, which would have helped her better imitate the child's positive affective expressions, is compromised.

We hereby link intersubjectivity and stress in terms of the dyadic active inference framework in three inter-related propositions, namely dyadic symbiosis, under-coupling, and over-mentalizing, as follows:

We hereby interpret MP intervention in light of the dyadic framework.

The three systems in the social brain, i.e., the mirroring system, mentalizing system, and salience network can be mapped to three components of the active inference framework. As depicted in Figure 1A, social cognition can be modeled as a hierarchical network of active inference engines, which encompasses: (1) an intermediate level involving mirroring system as a bottom-up component for automatic perception-action coupling, (2) the salience network as a feedback component mediating the surprise (i.e., socially salient prediction errors) detected in the intermediate level, and (3) a top level involving mentalizing system as a top-down component for affective and relational model to simulate relationships between self and others.

The functional distinction between the mirroring system and mentalizing system has gained empirical supports (112). As a bottom-up process, mirroring can be performed spontaneously without activating higher-order representations (16). In contrast, as a top-down process, while the mentalizing system can be activated by the theory-of-mind tasks (45), retrospective remembering and proactive imagining of episodic memory (113), and belief-based social attribution (46), it can also be spontaneously active without any inputs or task demands, as part of the default-mode network (61). The roles of the salience network in (a) conflict monitoring (49), (b) switching dynamic oscillations between the frontoparietal network (overlapping with the mirroring system) and the default-mode network (overlapping with the mentalizing system) during resting (66), and (c) representing signed prediction errors of reward (55, 56) and punishment (58, 59) are consistent with its potential role in the prediction errors as a feedback from the mirroring system to the mentalizing system.

This study advances the science of intersubjectivity and stress resilience on multiple levels. On a theory-generating level, we utilized a promising dyadic active inference framework and offered theoretical relationships between the “over-mentalizing” and “under-coupling” intersubjectivity problems and parenting stress. Further, on an empirical level, we proposed a novel fMRI task to identify neurocircuitry underlying intersubjectivity and potential mediators of the intersubjectivity-oriented intervention (Mom Power). Combined with the within-subject changes afforded by MP intervention, our results point to a two-pronged and potentially generalizable principle, i.e., stress resilience depends on not only mitigating stress-potentiated under-coupling and over-mentalizing problems, but also enhancing a stance of intersubjective benevolence in mirroring others' feelings and serving their well-being in dyadic symbiosis.