Hydrocephalus is a complex central nervous system (CNS) disorder characterized by an abnormal accumulation of cerebrospinal fluid (CSF). Although not a disease in itself, hydrocephalus manifests as a condition with diverse underlying causes. Congenital hydrocephalus exhibits a prevalence with rates of 50–70 per 100,000 births in North America, 80–110 in Europe, and up to 750 per 100,000 in Africa (JH, 2005; Warf, 2005; Munch et al., 2012; Dewan et al., 2018). The main treatment for hydrocephalus involves shunt surgery, a procedure designed to divert excess CSF and alleviate intracranial pressure (ICP) (Gholampour et al., 2019). Hydrocephalus is the leading reason for brain surgery in children, making it the most common condition requiring surgical intervention (Ponomaryov and Merkulova, 2015; Duy et al., 2023). Long-term evaluations have indicated that more than 85% of pediatric hydrocephalus patients may suffer shunt failure (Reddy et al., 2012). Infant patients show an even greater failure risk and lower shunt survival rates at all periods compared to older pediatric peers (McGirt et al., 2002; Shannon et al., 2012; Riva-Cambrin et al., 2016). Therefore, optimizing the functionality of the shunt system can be crucial to enhancing shunt efficiency and reducing shunt failure, particularly in the case of infant hydrocephalus.

Previous studies have attempted to consider the impact of various variables to reduce shunt failure and improve shunt outcomes. In this context, we evaluated the effect of different catheter routes on the risk of shunt failure (Gholampour et al., 2022b). Additionally, several studies have examined the impact of valve characteristics on shunt efficacy. Khan et al. and Gruber et al. provided compelling evidence on the substantial role of antisiphon devices of valves in reducing the complications of overdrainage following shunt surgery in infant and pediatric hydrocephalus patients (Gruber et al., 1984; Khan et al., 2010). Many studies have also compared the efficacy of various programmable and non-programmable shunt valves in pediatric and neonatal hydrocephalus, demonstrating a higher efficacy of programmable valves, specifically proGAV (McGirt et al., 2005; McGirt et al., 2007; Thomale et al., 2013; Beez et al., 2014; Hall et al., 2021b). It is worth mentioning that previous studies have shown that among these shunt valve characteristics, the valve drainage pressure setting, particularly the valve performance level, plays a crucial role in reducing the shunt failure rate and improving shunt outcomes in infant hydrocephalus (Hall et al., 2021a; Bock et al., 2021). The performance level of a valve in a hydrocephalus shunt pertains to the regulatory capacity of the shunt valve, allowing neurosurgeons to adjust the pressure or flow performance level. Adjusting the valve performance level presents a significant opportunity to control the amount of CSF to be withdrawn from the ventricular system and subarachnoid space. Furthermore, this adjustment can be effective in regulating the velocity of CSF drainage, thereby reducing shunt failure specifically by preventing overdrainage or underdrainage (Gölz et al., 2013). It can also control CSF drainage pressure, which, in turn, indirectly regulates ICP, the most important variable in the management of children’s hydrocephalus (Haridas and Tomita, 2017).

Table 1 presents comprehensive specifications of shunt valves, including manufacturers, models, types, pressure ranges, numbers and ranges of performance levels, and error margins of the most renowned manufacturers of shunt valves. Notably, there is significant variability in the adjustability of shunt valves, with differences observed in both the number (from 2 to 21) and range (from 1 to 6 cm H₂O) of performance levels across various valve models (Table 1). Yamada et al. indicated that a change in valve drainage pressure of 2–4 cm H₂O could optimize symptom improvement (Yamada et al., 2022). Conversely, Zemack et al. contended that a minimal valve adjustment of less than 2 cm H₂O could improve clinical outcomes; thus, considering a performance level range below 2 cm H₂O is rational to improve shunt outcomes (Zemack et al., 2003). To address these conflicts, efficient thresholds should be found to optimize the number and range of valve performance levels, based on their closer correlation with the improvement of shunt outcomes.

Medical machine learning has surfaced as a sophisticated approach to addressing classification, prediction, and optimization challenges, and it could prove useful in addressing our aforementioned concern. The comprehensive evaluation of valve characteristics requires concurrent consideration of the effects of all variables, such as anti-siphon mechanisms, programmable/non-programmable conditions, number of valve adjustments, and valve performance levels. Moreover, the multivariate nature of hydrocephalus gives rise to intricate and interrelated associations among these variables. Machine learning offers a unique opportunity to analyze the simultaneous impacts of all these variables, considering their complex interconnections. However, previous studies that have considered shunt valve characteristics in machine learning have been significantly scarce (Hale et al., 2021). This study seeks to bridge the research gap related to optimizing and standardizing valve performance level information. The goals of the present study are to establish the optimized number of shunt valve performance levels and to delineate the optimized valve drainage pressure ranges corresponding to each performance level through unsupervised machine learning analysis. This study is designed to utilize its findings to potentially improve the efficacy of shunt performance in treating hydrocephalus patients who were under 1 year of age at the time of their initial shunt surgery.

The study design, protocols, and procedures were approved by the Human Institutional Review Board (IRB) committee and adhered to the ethics guidelines of the University of Chicago (IRB code: IRB23-1157), based on the 1964 Helsinki Declaration and its later amendments.

The present study only focused on infant hydrocephalus patients with first shunt surgery without any failure or complications to find the best clustering based on shunt surgeries with improved outcomes. An initial cohort of 148 hydrocephalus patients aged less than 1 year with a history of shunting was identified (Figure 1). 97 patients were excluded from analysis due to subsequent shunt infection, revision, failure, unimproved outcomes, or incomplete data. This resulted in a final cohort of 51 infant hydrocephalus patients who underwent shunt surgery without subsequent failure or unimproved outcomes (Figure 1). Table 2 indicates the baseline characteristics of the infant hydrocephalus patient cohort used in the present study. The shunt failure rate for the initial shunt surgery of our patients was 65.3%. To understand the age demographics of our patient cohort, we utilized an empirical cumulative distribution function plot (Figure 2). This plot provides a visual representation of the cumulative distribution of ages among the patients. Remarkably, the analysis revealed a pivotal insight: 50% of our patients were aged less than 3.3 months. This finding underscores the significance of early-age patients within our study, shedding light on a critical demographic aspect that might influence various aspects of our analysis, from disease prevalence to treatment strategies. To verify improvements in clinical outcomes post-shunt surgery, patient documents were monitored on EPIC, with an average follow-up period of 94.9 months subsequent to the initial shunt procedure, as indicated in Table 2.

A three-dimensional principal component analysis (PCA) was performed on our dataset of patients and their corresponding features (Figure 3). This analysis reduced the dataset’s dimensionality while preserving variance, allowing for a nuanced exploration of complex patient data. Patients are depicted as points in a 3D space, with their positions determined by their combined features.

One of the most significant considerations in clustering is determining the optimal number of clusters. The Silhouette coefficient, which measures cluster cohesion and separation, provides valuable insights for this determination. The clustering solution with 4 clusters demonstrated a remarkable performance, exhibiting a Silhouette coefficient that was at least 28.3% higher than that of the solutions with 3 and 5-7 clusters (Figure 4A). This higher Silhouette coefficient indicates that the individual data within each cluster were, on average, more similar to each other and more dissimilar to individual data in other clusters in the 4-cluster solution compared to the other considered cluster formations. Furthermore, the DB index, which evaluates the compactness and separation between clusters, also provided additional support for the superiority of the 4-cluster solution. The DB index for the 4-cluster solution was at least 37.2% lower than that of the solutions with 3 and 5-7 clusters (Figure 4B). This lower index value signifies that the clusters in the 4-cluster solution were more well-defined and distinct from each other, reinforcing the efficacy of this particular clustering solution. Therefore, both the Silhouette coefficient and DB index confirmed that the 4-cluster configuration is the optimal choice for clustering.

The resultant dendrogram, as depicted in Figure 5, provides a visual representation of the hierarchical clustering process, illustrating not only the arrangement of the data clusters but also the sequence and distance at which individual data points are merged into clusters. From this dendrogram analysis, we extracted four robust clusters (Figure 5: red dash line), labeled as Cluster 1 (valve performance level 1), Cluster 2 (valve performance level 2), Cluster 3 (valve performance level 3), and Cluster 4 (valve performance level 4). Each cluster encapsulates unique subsets of our data without any overlaps between the clusters. To enhance the interpretability of our clustering results, we generated a scatter plot (Figure 6). Each data point is represented, with distinct clusters color-coded. The plot effectively demonstrates the hierarchical clustering method’s ability to categorize the data into four clusters. Figure 7 showcases the distance matrix heatmap, which is central to elucidating the complex relationships among our patient cases based on pairwise dissimilarities. In this matrix, lighter shades indicate closer relationships, suggesting smaller distances between patient cases, whereas darker shades signify greater dissimilarities or larger distances. This matrix summarizes the hierarchical clustering iterations, detailing merged clusters, distances, and new cluster sizes. Based on the results from Figures 5–7, we organized and arranged the valve drainage pressure values for patients in each cluster in Figure 8. Box plots in Figure 8A highlight pressure differences across performance levels (levels 1–4). Each box represents the interquartile range, with the median shown inside, and whiskers display the pressure variability within clusters.

Previous research has highlighted the significant impact of optimizing valve performance levels, especially during the initial adjustment following the first shunt surgery, on shunt efficacy and the improvement of outcomes for infant hydrocephalus (Hall et al., 2021a; Bock et al., 2021). However, there are conflicts and contradictions about the optimal pressure range for each valve performance level, and consequently, the number of valve performance levels (Zemack et al., 2003; Yamada et al., 2022), among the large number of performance levels (ranging from 2 to 21) and a wide pressure range within each level (1–6 cm H₂O) (Table 1). Therefore, finding the optimal thresholds for valve performance levels, which correlate more closely with the improvement of shunt outcomes, is of great importance. The present study aimed to optimize the number and range of valve performance levels to potentially assist neurosurgeons in enhancing shunt performance and improving clinical outcomes for infant hydrocephalus.

Our findings demonstrate that, in terms of both intra-cluster similarity (indicated by the Silhouette coefficient) and inter-cluster dissimilarity (measured by the DB index), four valve performance levels outperform other alternatives (Figure 4). The results presented in Figure 4 also indicate that a three-level valve emerges as the next best alternative, as measured by intra-cluster similarity (Silhouette coefficient) and inter-cluster dissimilarity (DB index). However, practical and clinical considerations led us to focus solely on the first option (four-level valve). From a clinical and neurosurgical perspective, a three-level system is somewhat unconventional. To the best of our knowledge, none of the current shunts on the market utilizes a three-level system, which suggests a potential limitation in its functionality and its acceptance in current medical practice. Notably, the calculated valve drainage pressure ranges for each performance level were completely separated and distinct, and it was crucial for us to have clustering without any overlaps between different performance levels (Figures 6, 8). The repeatability of our results is essential for the implementation of our findings in a clinical study. Therefore, in addition to the acceptable results obtained from the Silhouette coefficient and the DB index for four clusters, cluster stability analysis becomes crucial for evaluating the repeatability, robustness, and reliability of our clustering results, particularly for relatively small data sizes. Employing bootstrapping techniques, we generated numerous iterations of resampled datasets, equal in size to the original dataset, to enable a rigorous and extensive assessment of the stability across our clustering analysis iterations. We employed the Jaccard index as our stability metric, enabling a precise comparison of clustering results from different resampled datasets. Our analysis revealed a notable Jaccard index of 0.71 for the 4-cluster, underscoring the reliability and repeatability of our chosen cluster and indicating that the identified clusters were not significantly influenced by random variations (Figure 9).

The cluster profile, as detailed in Table 3, encapsulates the diversity of our patient cohort across each cluster, distilling their multifaceted characteristics into four distinctive columns. Each row in this table represents a unique patient profile or feature. As per the results presented in Table 3, the mean (standard deviation) values of valve drainage pressure for performance levels 1, 2, 3, and 4 are 3.6 (0.86), 7.5 (0.62), 12.2 (0.99), and 16.7 (0.73) cm H₂O, respectively. Understanding these distinct profiles is crucial as it enables clinicians and researchers to identify commonalities among patients within each cluster as patients within the same cluster might respond similarly to specific interventions. Figure 8B displays our proposed valve performance settings for infant hydrocephalus, featuring minimum and maximum values for the four performance levels (pressure range): 1.5–5 (3.5) cm H₂O, 5–9 (4) cm H₂O, 9–15 (6) cm H₂O, and 15–18 (3) cm H₂O, respectively.

Some widely-used valve brands, including Medtronic-Delta, also exhibit four valve performance levels, similar to our result. In the majority of common valve models such as Medtronic (Strata II, Strata MR II, Delta, and CSF Flow Control), Integra (Certas Plus), Codman (Uni-Shunt), Integra (Integra DP and Atlas), Natus (Novus), and Radionics (Equiflow), the pressure ranges in performance levels are not equal (Table 1), aligning with our finding. This finding may be attributed to the inherent variability in shunt outcomes among infant patients, where a consistent impact across equal pressure ranges was not always evident. On the other hand, certain shunt valve models, such as the Medtronic Strata II, Strata MR II, and Delta, as well as the Integra Certas Plus, define their valve drainage pressure ranges as less than or equal to 1 cm H₂O per performance level (Table 1). In contrast, our findings—derived from patients without shunt failure and with improved shunt outcomes—revealed a broader pressure range of at least 3 cm H₂O for each performance level. Although this differs from the narrower ranges set by these companies, it aligns more closely with the majority of other valve types (Table 1).

In our investigation of the etiology of infant hydrocephalus, we categorized etiological conditions in relation to different valve levels, as presented in rows 3-8 of Table 3. While these classifications yielded initial insights, they did not demonstrate a definitive separation of etiology across different valve levels, reflecting the inherent complexity of hydrocephalus etiology. This observation suggests that the relationship between specific valve settings and the underlying causes of hydrocephalus is not straightforward, indicating the need for more nuanced research approaches.

Our recent studies have emphasized the significance of time in the process of reducing ICP and enhancing intracranial compliance, and its impact on the improvement of clinical outcomes in gradual-onset brain disorders such as hydrocephalus (Gholampour, 2023a; Gholampour, 2023b). We have also highlighted the critical importance of precise control over the timing of CSF drainage following shunt surgery (Gholampour et al., 2022a; Gholampour et al., 2022c). We demonstrated that this control significantly influences the recovery behavior of brain tissue, thereby enhancing the efficiency of shunt performance and improving patient outcomes (Gholampour et al., 2022a). Our proposed optimization of the number and range of valve performance levels may also prove beneficial in enhancing control over the timing of CSF drainage. This, in turn, could facilitate the appropriate adaptation of brain tissue, ensuring proper brain recovery and increasing the likelihood of improved shunt outcomes.

This study strove to provide neurosurgeons with data-driven information about the optimized number and range of valve performance levels that may improve surgical outcomes in infant hydrocephalus. While our findings may offer valuable insights, it is important to acknowledge some limitations that should be addressed in future research. The specific geometry of shunt valves holds notable potential to influence their fluid resistance, subsequently impacting drainage flow rates within established pressure ranges. However, some literature suggests that this influence might not play a significant role in determining shunt outcomes (Hoshide et al., 2017). It is essential to recognize that while a linear relationship exists between the pressure and flow rate of CSF drainage, this relationship can vary between different valves, exhibiting diverse slopes. In our study, we have incorporated valve drainage pressure as one of the features in our machine learning analysis, thus indirectly accounting for its implication on flow rate. However, it is suggested for future studies to investigate the effects of various valve geometries on our optimization results. It is also important to highlight that this study specifically focused on pressure-controlled valves. However, we acknowledge the potential benefits of self-adjustable valves and recommend replicating our findings for self-adjustable valves in future research. Additionally, while our retrospective machine learning analysis is grounded in clinical data, it is important to note that this was a retrospective study. We recommend that future research adopt a prospective approach to apply our findings in a clinical trial with larger patient cohorts. This approach will facilitate the validation of the effectiveness of the optimization results and enable the assessment of the correlation between our proposed valve levels and the varying levels of clinical outcomes in both infant and older pediatric patients with hydrocephalus across diverse clinical scenarios. Moreover, this was an unsupervised (non-classification) machine learning analysis; therefore, focusing on patients with unimproved outcomes in future studies could be beneficial for conducting error analysis.

In this study, we applied unsupervised hierarchical machine learning analysis to cluster infant hydrocephalus patients based on features such as age, gender, hydrocephalus etiology (six categories), type of hydrocephalus (two categories), shunt catheter route (three categories), anti-siphon condition, programmable status, number of valve adjustments, and valve drainage pressure. The evaluation of the Silhouette coefficient, DB index, and cluster stability analysis using the Jaccard index unequivocally established that the optimal number of valve performance levels for infant hydrocephalus was four: 1.5–5.0, 5.0–9.0, 9.0–15.0, and 15.0–18.0 cm H₂O. The recommendation of our optimized number and range of valve performance levels holds the potential for neurosurgeons to improve shunt outcomes in infant hydrocephalus, ultimately enhancing the quality of life for these patients.