I have to agree with @Thugmek, you need to experiment your neural architecture. I'm going to put some code to his suggestion:

Thugnek's Suggestion: Add layers plus a few modifications to experiment with

    // add layers to the model
    model.add(tf.layers.dense({ units: 64, activation: 'relu', inputShape: [6] }));
    model.add(tf.layers.dropout({ rate: 0.2 })); // Add dropout to prevent overfitting
    model.add(tf.layers.dense({ units: 32, activation: 'relu' }));
    model.add(tf.layers.dropout({ rate: 0.2 })); // Add dropout to prevent overfitting

    // softmax because multiclass
    model.add(tf.layers.dense({ units: uniqLabels.length, activation: 'softmax' }));

    // try a better optimizer like adam
    model.compile({
        optimizer: 'adam', // Adam is a robust optimizer over SGD for deep learning
        loss: 'sparseCategoricalCrossentropy',
        metrics: ['accuracy']
    });